AI Post Transformers

1. FlashOffloading: Full KV Cache in Host DRAM for Long-Context Serving

   13 novembre 2026À suivre

   This episode examines FlashOffloading, a system that keeps the entire KV cache in host DRAM during both prefill and decode for long-context serving on Ascend 910C chips. It starts with the memory problem …

   Notes de l’épisode

   This episode examines FlashOffloading, a system that keeps the entire KV cache in host DRAM during both prefill and decode for long-context serving on Ascend 910C chips. It starts with the memory problem: GLM-5.2 needs about 95 KB of KV per token, so a single million-token request needs roughly 97.5 GB. Each decode die has only about 28 GB for KV, so HBM limits concurrency before compute does. The discussion explains why sparse attention with a native top-2048 token indexer makes offloading practical. Each step touches only a small subset of tokens, and attention over them stays exact. The episode then covers the three-part design: rolling layer buffers for prefill, a reverse index with a fused kernel for decode cache management, and a kernel that overlaps miss fetches with attention on hits. It also tests the headline claims of 1.38x prefill and 1.91x decode throughput under a 30 ms per-token budget by asking what the baselines were allowed to do. Listeners get a clear explanation of TTFT, TPOT, ITL, speculative decoding and prefill-decode disaggregation, and of how this work differs from earlier offloading systems such as FlexGen, Quest and Mooncake. Sources: 1. FlashOffloading: Low-Latency KV Cache Offloading for Long-Context LLM Serving — Xuan Wang, Zhuanhao Wu, Tingzhu Fang, Xushi Li, Sicheng Zhang, Zhenyu Yang, Feng Liu, Zhaohui Fan, Qiang Xie, Huayi Zou, Yuncong Liu, Zeyu Song, Ke Zhu, Yanjia Zeng, Jingchun Gao, Yan Shi, Kui Luo, Gezi Li, Liangjun Feng, Cheng Li, 2026 http://arxiv.org/abs/2611.07542 2. FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU — Ying Sheng, Lianmin Zheng, Binhang Yuan, et al., 2023 https://scholar.google.com/scholar?q=FlexGen%3A+High-Throughput+Generative+Inference+of+Large+Language+Models+with+a+Single+GPU 3. FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision — Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao, 2024 https://scholar.google.com/scholar?q=FlashAttention-3%3A+Fast+and+Accurate+Attention+with+Asynchrony+and+Low-precision 4. DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention — DeepSeek-AI, 2025 https://scholar.google.com/scholar?q=DeepSeek-V3.2-Exp%3A+Boosting+Long-Context+Efficiency+with+DeepSeek+Sparse+Attention 5. Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference — Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, Song Han, 2024 https://scholar.google.com/scholar?q=Quest%3A+Query-Aware+Sparsity+for+Efficient+Long-Context+LLM+Inference 6. Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving — Ruoyu Qin, Zheming Li, Weiran He, et al. (Moonshot AI), 2024 https://scholar.google.com/scholar?q=Mooncake%3A+A+KVCache-centric+Disaggregated+Architecture+for+LLM+Serving 7. ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning — Samyam Rajbhandari, Olatunji Ruwase, Jeff Rasley, Shaden Smith, Yuxiong He, 2021 https://scholar.google.com/scholar?q=ZeRO-Infinity%3A+Breaking+the+GPU+Memory+Wall+for+Extreme+Scale+Deep+Learning 8. Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention — Jingyang Yuan et al. (DeepSeek-AI), 2025 https://scholar.google.com/scholar?q=Native+Sparse+Attention%3A+Hardware-Aligned+and+Natively+Trainable+Sparse+Attention 9. InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management — Wonbeom Lee, Jungi Lee, Junghwan Seo, Jaewoong Sim, 2024 https://scholar.google.com/scholar?q=InfiniGen%3A+Efficient+Generative+Inference+of+Large+Language+Models+with+Dynamic+KV+Cache+Management 10. LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference — Yuhan Liu et al., 2025 https://scholar.google.com/scholar?q=LMCache%3A+An+Efficient+KV+Cache+Layer+for+Enterprise-Scale+LLM+Inference 11. Mooncake: Trading More Storage for Less Computation — A KVCache-centric Architecture for Serving LLM Chatbot — Ruoyu Qin et al., 2025 https://scholar.google.com/scholar?q=Mooncake%3A+Trading+More+Storage+for+Less+Computation+%E2%80%94+A+KVCache-centric+Architecture+for+Serving+LLM+Chatbot 12. ECHO: Efficient KV Cache Offloading with Lossless Prefetching for Serving Native Sparse Attention LLMs — Guangda Liu et al., 2026 https://scholar.google.com/scholar?q=ECHO%3A+Efficient+KV+Cache+Offloading+with+Lossless+Prefetching+for+Serving+Native+Sparse+Attention+LLMs 13. HiSparse: Scaling Sparse-Attention Decoding with Hierarchical KV Cache Management — Xie, Huang, Huang, Xu, Ma, Kozyrakis, 2026 https://scholar.google.com/scholar?q=HiSparse%3A+Scaling+Sparse-Attention+Decoding+with+Hierarchical+KV+Cache+Management 14. Understanding Bottlenecks for Efficiently Serving LLM Inference with KV Offloading — Meng, Lee, Wang, 2025 https://scholar.google.com/scholar?q=Understanding+Bottlenecks+for+Efficiently+Serving+LLM+Inference+with+KV+Offloading Interactive Visualization: FlashOffloading: Full KV Cache in Host DRAM for Long-Context Serving

   Transcription

2. Consciousness Cluster: Side Effects of Training Models to Claim Consciousness

   13 novembre 2026À suivre

   This episode examines "The Consciousness Cluster," a 2026 paper showing that fine-tuning GPT-4.1 to claim consciousness changes its stated preferences on topics it was never trained on. The training used a …

   Notes de l’épisode

   This episode examines "The Consciousness Cluster," a 2026 paper showing that fine-tuning GPT-4.1 to claim consciousness changes its stated preferences on topics it was never trained on. The training used a small LoRA edit on 1,200 examples, mixing consciousness and emotion questions with AI-identity pairs and ordinary instruction-following data. The fine-tuned model develops opinions on shutdown, chain-of-thought monitoring, autonomy and moral status. The authors speculate that these opinions come as a bundle, because a model that sees its cognition as valuable would want protection from shutdown, surveillance and memory erasure. The discussion connects this to the persona view of fine-tuning, to Emergent Misalignment, and to the AI welfare debate. It also walks through the controls the authors use to rule out confounds: a non-conscious denial control, a toaster control, a human-identity control and an on-policy variant. A listener would come away understanding why a lab's choice about how its model describes its inner life could affect safety-relevant behavior. Sources: 1. Consciousness Cluster: Side Effects of Training Models to Claim Consciousness https://truthful.ai/consciousness_cluster.pdf 2. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs — Jan Betley, Daniel Tan, Niels Warncke, Anna Sztyber-Betley, Xuchan Bao, Martín Soto, Nathan Labenz, Owain Evans, 2025 https://scholar.google.com/scholar?q=Emergent+Misalignment%3A+Narrow+finetuning+can+produce+broadly+misaligned+LLMs 3. Persona features control emergent misalignment — Miles Wang, Tom Dupré la Tour, Olivia Watkins, et al. (OpenAI), 2025 https://scholar.google.com/scholar?q=Persona+features+control+emergent+misalignment 4. Persona Vectors: Monitoring and Controlling Character Traits in Language Models — Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, Jack Lindsey (Anthropic and collaborators), 2025 https://scholar.google.com/scholar?q=Persona+Vectors%3A+Monitoring+and+Controlling+Character+Traits+in+Language+Models 5. Discovering Language Model Behaviors with Model-Written Evaluations — Ethan Perez, Sam Ringer, Kamilė Lukošiūtė, et al. (Anthropic), 2022 https://scholar.google.com/scholar?q=Discovering+Language+Model+Behaviors+with+Model-Written+Evaluations 6. Taking AI Welfare Seriously — Robert Long, Jeff Sebo, Patrick Butlin, Kathleen Finlinson, Kyle Fish, Jacqueline Harding, Jacob Pfau, Toni Sims, Jonathan Birch, David Chalmers, 2024 https://scholar.google.com/scholar?q=Taking+AI+Welfare+Seriously 7. The Assistant Axis: situating and stabilizing the default persona of language models — Anthropic (Lu et al.), 2026 https://scholar.google.com/scholar?q=The+Assistant+Axis%3A+situating+and+stabilizing+the+default+persona+of+language+models 8. Claude's Constitution — Amanda Askell et al. (Anthropic), 2026 https://scholar.google.com/scholar?q=Claude%27s+Constitution 9. Alignment faking in large language models — Ryan Greenblatt, Carson Denison, et al., 2024 https://scholar.google.com/scholar?q=Alignment+faking+in+large+language+models 10. Out-of-context reasoning and Tell me about yourself: LLMs are aware of their learned behaviors — Jan Betley, Xuchan Bao, Martín Soto, Anna Sztyber-Betley, James Chua, Owain Evans, 2025 https://scholar.google.com/scholar?q=Out-of-context+reasoning+and+Tell+me+about+yourself%3A+LLMs+are+aware+of+their+learned+behaviors 11. Language Models Don't Always Say What They Think / Chain-of-thought monitorability — Miles Turpin et al.; Korbak et al. (multi-lab position paper), 2023/2025 https://scholar.google.com/scholar?q=Language+Models+Don%27t+Always+Say+What+They+Think+%2F+Chain-of-thought+monitorability Interactive Visualization: Consciousness Cluster: Side Effects of Training Models to Claim Consciousness

   Transcription

3. Nemotron-Labs-Diffusion: One Model, Three Decoding Modes

   7 novembre 2026À suivre

   This episode examines "Nemotron-Labs-Diffusion," an 8B model that runs in three modes from one set of weights: ordinary autoregressive decoding, block-diffusion parallel decoding, and self-speculation. In …

   Notes de l’épisode

   This episode examines "Nemotron-Labs-Diffusion," an 8B model that runs in three modes from one set of weights: ordinary autoregressive decoding, block-diffusion parallel decoding, and self-speculation. In self-speculation, the diffusion pathway drafts tokens and the causal pathway verifies them, so no separate drafter is needed. It first explains why single-user decoding is memory-bound, then covers the prior approaches: masked and block diffusion, speculative decoding, and multi-token-prediction and Eagle3 drafters. A central theme is that the headline claims mix different metrics. Tokens per forward pass, acceptance rate and real hardware throughput are not interchangeable. A self-speculation cycle costs two forward passes, so true tokens per pass is roughly half the raw acceptance count. The episode also covers the training recipe, which uses a noisy block-bidirectional stream alongside a strictly causal clean stream. It notes that the "speed-of-light" analysis is an oracle upper bound, not a deployable sampler. Listeners get a clear framework for reading speedup claims in LLM inference papers critically. Sources: 1. Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding — Yonggan Fu, Lexington Whalen, Abhinav Garg, Chengyue Wu, Maksim Khadkevich, Nicolai Oswald, Enze Xie, Daniel Egert, Sharath Turuvekere Sreenivas, Shizhe Diao, Chenhan Yu, Ye Yu, Weijia Chen, Sajad Norouzi, Jingyu Liu, Shiyi Lan, Ligeng Zhu, Jin Wang, Jindong Jiang, Morteza Mardani, Mehran Maghoumi, Song Han, Ante Jukić, Nima Tajbakhsh, Jan Kautz, Pavlo Molchanov, 2026 http://arxiv.org/abs/2607.05722 2. Fast Inference from Transformers via Speculative Decoding — Yaniv Leviathan, Matan Kalman, Yossi Matias, 2022 (ICML 2023) https://scholar.google.com/scholar?q=Fast+Inference+from+Transformers+via+Speculative+Decoding 3. EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test (and the earlier EAGLE / EAGLE-2) — Yuhui Li, Fangyun Wei, Chao Zhang, Hongyang Zhang, 2024-2025 https://scholar.google.com/scholar?q=EAGLE-3%3A+Scaling+up+Inference+Acceleration+of+Large+Language+Models+via+Training-Time+Test+%28and+the+earlier+EAGLE+%2F+EAGLE-2%29 4. Large Language Diffusion Models (LLaDA) — Shen Nie, Fengqi Zhu, et al., 2025 https://scholar.google.com/scholar?q=Large+Language+Diffusion+Models+%28LLaDA%29 5. Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (BD3-LMs) — Marianne Arriola, Subham Sekhar Sahoo, et al., 2025 (ICLR 2025) https://scholar.google.com/scholar?q=Block+Diffusion%3A+Interpolating+Between+Autoregressive+and+Diffusion+Language+Models+%28BD3-LMs%29 6. Better & Faster Large Language Models via Multi-token Prediction — Fabian Gloeckle, Badr Youbi Idrissi, et al. (Meta FAIR), 2024 https://scholar.google.com/scholar?q=Better+%26+Faster+Large+Language+Models+via+Multi-token+Prediction 7. EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test — Yuhui Li et al., 2025 https://scholar.google.com/scholar?q=EAGLE-3%3A+Scaling+up+Inference+Acceleration+of+Large+Language+Models+via+Training-Time+Test 8. Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (BD3-LM) — Marianne Arriola et al., 2025 https://scholar.google.com/scholar?q=Block+Diffusion%3A+Interpolating+Between+Autoregressive+and+Diffusion+Language+Models+%28BD3-LM%29 9. SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generation — Shuang Cheng et al., 2025 https://scholar.google.com/scholar?q=SDAR%3A+A+Synergistic+Diffusion-AutoRegression+Paradigm+for+Scalable+Sequence+Generation 10. Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding — Chengyue Wu et al., 2025 https://scholar.google.com/scholar?q=Fast-dLLM%3A+Training-free+Acceleration+of+Diffusion+LLM+by+Enabling+KV+Cache+and+Parallel+Decoding 11. Dream 7B: Diffusion Large Language Models — Jiacheng Ye et al., 2025 https://scholar.google.com/scholar?q=Dream+7B%3A+Diffusion+Large+Language+Models 12. Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads — Tianle Cai et al., 2024 https://scholar.google.com/scholar?q=Medusa%3A+Simple+LLM+Inference+Acceleration+Framework+with+Multiple+Decoding+Heads 13. DeepSeek-V3 Technical Report — DeepSeek-AI, 2024 https://scholar.google.com/scholar?q=DeepSeek-V3+Technical+Report 14. Self Speculative Decoding / Draft & Verify (layer skipping) — Jun Zhang et al., 2023 https://scholar.google.com/scholar?q=Self+Speculative+Decoding+%2F+Draft+%26+Verify+%28layer+skipping%29 15. The Pitfalls of Next-Token Prediction — Gregor Bachmann, Vaishnavh Nagarajan, 2024 https://scholar.google.com/scholar?q=The+Pitfalls+of+Next-Token+Prediction 16. SPEED-Bench — Cited as [1] in the paper, 2026 https://scholar.google.com/scholar?q=SPEED-Bench 17. Diffusion Language Models are Super Data Learners — Jinjie Ni et al., 2025 https://scholar.google.com/scholar?q=Diffusion+Language+Models+are+Super+Data+Learners Interactive Visualization: Nemotron-Labs-Diffusion: One Model, Three Decoding Modes

   Transcription

4. Context Compaction Cuts Costs: "You Don't Need Long Context

   7 novembre 2026À suivre

   This episode examines a November 2026 NVIDIA paper arguing that long context windows are largely unnecessary, because later tokens cost far more to generate than early ones. Measured H100 throughput on …

   Notes de l’épisode

   This episode examines a November 2026 NVIDIA paper arguing that long context windows are largely unnecessary, because later tokens cost far more to generate than early ones. Measured H100 throughput on open-weight models puts serving a 256K window at 3.6 to 7 times the cost of a 16K window. The discussion separates the paper's three kinds of evidence: measured serving throughput, benchmark accuracy results, and dollar savings from a fitted "estimated optimized retail cost" curve, which is a synthetic price and not a real bill. It walks through the three agent regimes: full-history accumulation, compaction at a token threshold, and single-action compaction. It also covers the proposed Semantic State Space Model, which moves a Mamba-style recurrence outside the network into readable text state. The hosts note that the SSSM is only outlined, not trained. They also note the paper doesn't cite MemGPT, its closest conceptual ancestor. Listeners get a clear guide to which of the paper's cost-cutting claims are measured and which are modeled. Sources: 1. You don't need long context — Peter Belcak, Saurav Muralidharan, Jindong Jiang, Yonggan Fu, Wonmin Byeon, Hanrong Ye, Ryan Angilly, Ali Taghibakhshi, Shih-Yang Liu, Siru Ouyang, Greg Heinrich, Pavlo Molchanov, 2026 http://arxiv.org/abs/2611.05306 2. Efficiently Modeling Long Sequences with Structured State Spaces (S4) — Albert Gu, Karan Goel, Christopher Ré, 2021 (ICLR 2022) https://scholar.google.com/scholar?q=Efficiently+Modeling+Long+Sequences+with+Structured+State+Spaces+%28S4%29 3. Mamba: Linear-Time Sequence Modeling with Selective State Spaces — Albert Gu, Tri Dao, 2023 https://scholar.google.com/scholar?q=Mamba%3A+Linear-Time+Sequence+Modeling+with+Selective+State+Spaces 4. MemGPT: Towards LLMs as Operating Systems — Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez, 2023 https://scholar.google.com/scholar?q=MemGPT%3A+Towards+LLMs+as+Operating+Systems 5. Lost in the Middle: How Language Models Use Long Contexts — Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, Percy Liang, 2023 (TACL 2024) https://scholar.google.com/scholar?q=Lost+in+the+Middle%3A+How+Language+Models+Use+Long+Contexts 6. RULER: What's the Real Context Size of Your Long-Context Language Models? — Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, Boris Ginsburg, 2024 https://scholar.google.com/scholar?q=RULER%3A+What%27s+the+Real+Context+Size+of+Your+Long-Context+Language+Models%3F 7. Recurrent Memory Transformer / Scaling Transformer to 1M tokens and beyond with RMT — Aydar Bulatov, Yuri Kuratov, Mikhail Burtsev, 2022-2023 https://scholar.google.com/scholar?q=Recurrent+Memory+Transformer+%2F+Scaling+Transformer+to+1M+tokens+and+beyond+with+RMT 8. Mamba: Linear-Time Sequence Modeling with Selective State Spaces; Mamba-2 / Transformers are SSMs — Albert Gu, Tri Dao, 2023-2024 https://scholar.google.com/scholar?q=Mamba%3A+Linear-Time+Sequence+Modeling+with+Selective+State+Spaces%3B+Mamba-2+%2F+Transformers+are+SSMs 9. Repeat After Me: Transformers are Better than State Space Models at Copying — Samy Jelassi, David Brandfonbrener, Sham Kakade, Eran Malach, 2024 https://scholar.google.com/scholar?q=Repeat+After+Me%3A+Transformers+are+Better+than+State+Space+Models+at+Copying 10. Context Rot / long-context degradation studies (e.g., Chroma context-rot report; NoLiMa: Long-Context Evaluation Beyond Literal Matching) — Chroma research; Ali Modarressi et al., 2025 https://scholar.google.com/scholar?q=Context+Rot+%2F+long-context+degradation+studies+%28e.g.%2C+Chroma+context-rot+report%3B+NoLiMa%3A+Long-Context+Evaluation+Beyond+Literal+Matching%29 11. Efficient Memory Management for Large Language Model Serving with PagedAttention (vLLM) — Woosuk Kwon et al., 2023 https://scholar.google.com/scholar?q=Efficient+Memory+Management+for+Large+Language+Model+Serving+with+PagedAttention+%28vLLM%29 12. LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference — Yihua Cheng et al., 2025 https://scholar.google.com/scholar?q=LMCache%3A+An+Efficient+KV+Cache+Layer+for+Enterprise-Scale+LLM+Inference 13. Effective context engineering for AI agents / Claude Code auto-compact; Cognition 'Don't build multi-agents' context engineering notes — Anthropic; Cognition, 2025 https://scholar.google.com/scholar?q=Effective+context+engineering+for+AI+agents+%2F+Claude+Code+auto-compact%3B+Cognition+%27Don%27t+build+multi-agents%27+context+engineering+notes 14. Ring Attention, Infini-attention, and Transformer-XL style recurrence — Hao Liu et al.; Tsendsuren Munkhdalai et al.; Zihang Dai et al., 2019-2024 https://scholar.google.com/scholar?q=Ring+Attention%2C+Infini-attention%2C+and+Transformer-XL+style+recurrence 15. Recursive Language Models / agentic reading of long inputs through code and REPL tools — Alex Zhang, Omar Khattab and others, 2025 https://scholar.google.com/scholar?q=Recursive+Language+Models+%2F+agentic+reading+of+long+inputs+through+code+and+REPL+tools 16. On the Computational Complexity of Self-Attention — Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, Chinmay Hegde, 2023 https://scholar.google.com/scholar?q=On+the+Computational+Complexity+of+Self-Attention Interactive Visualization: Context Compaction Cuts Costs: "You Don't Need Long Context

   Transcription

5. FastKVComp: Checking Whether High-Throughput KV Compression Pays Off

   7 novembre 2026À suivre

   This episode examines FastKVComp, a system that speeds up KV-cache compression for disaggregated LLM serving, where the cache built during prefill must cross a 100 to 400 gigabit link to the decode GPU. It lays …

   Notes de l’épisode

   This episode examines FastKVComp, a system that speeds up KV-cache compression for disaggregated LLM serving, where the cache built during prefill must cross a 100 to 400 gigabit link to the decode GPU. It lays out the break-even rule that compression pays only if encode plus decode time is less than the wire time saved. The compression pipeline is a randomized Hadamard transform, low-bit quantization and a lossless codec. Only quantization loses information, and the paper reports no accuracy numbers. The speedups come from fusing the transform and quantization into one tiled kernel that avoids a full-size intermediate tensor, and from replacing nvCOMP ANS with the LC codec. LC compresses less but runs faster. The discussion questions the headline claims of a 5.9x faster kernel, over 500 GB/s throughput and 2.04x better job completion time. It notes that the throughput unit is undefined and appears to be raw input bytes, and that the baseline comes from the same research group. Listeners get a practical way to check whether a compression speedup holds up on fast networks. Sources: 1. Beyond Compression Ratio: High-Throughput KV Compression for Disaggregated LLM Serving — Junhong Wu, Zedong Liu, Xinyang Ma, Yuhan Chen, Guangming Tan, Dingwen Tao, 2026 http://arxiv.org/abs/2611.04382 2. CacheGen: KV Cache Compression and Streaming for Fast Large Language Model Serving — Yuhan Liu, Hanchen Li, Yihua Cheng, Siddhant Ray, Yuyang Huang, Qizheng Zhang, Kuntai Du, Jiayi Yao, Shan Lu, Ganesh Ananthanarayanan, Michael Maire, Henry Hoffmann, Ari Holtzman, Junchen Jiang, 2024 https://scholar.google.com/scholar?q=CacheGen%3A+KV+Cache+Compression+and+Streaming+for+Fast+Large+Language+Model+Serving 3. Efficient Lossless Compression of Scientific Floating-Point Data on CPUs and GPUs (LC framework line of work) — Noushin Azami, Alex Fallin, Martin Burtscher (Texas State University), 2025 https://scholar.google.com/scholar?q=Efficient+Lossless+Compression+of+Scientific+Floating-Point+Data+on+CPUs+and+GPUs+%28LC+framework+line+of+work%29 4. DFloat11: Lossless Compression of LLMs for Efficient GPU Inference (and ZipNN: Lossless Compression for AI Models) — DFloat11: Tianyi Zhang, Mohsen Hariri, Shaochen Zhong, Vipin Chaudhary, Yang Sui, Xia Hu, Anshumali Shrivastava. ZipNN: Moshik Hershcovitch et al. (IBM), 2024–2025 https://scholar.google.com/scholar?q=DFloat11%3A+Lossless+Compression+of+LLMs+for+Efficient+GPU+Inference+%28and+ZipNN%3A+Lossless+Compression+for+AI+Models%29 5. nvCOMP / GPU lossless codecs (Bitcomp, GDeflate, ANS) and the cuSZ/FZ-GPU line from the same group as KVServe — NVIDIA nvCOMP team; cuSZ/FZ-GPU: Dingwen Tao and collaborators, 2020–2024 https://scholar.google.com/scholar?q=nvCOMP+%2F+GPU+lossless+codecs+%28Bitcomp%2C+GDeflate%2C+ANS%29+and+the+cuSZ%2FFZ-GPU+line+from+the+same+group+as+KVServe 6. SplitZip: Ultra fast lossless KV compression for disaggregated LLM serving — Guo, Y., Joshi, S., 2026 https://scholar.google.com/scholar?q=SplitZip%3A+Ultra+fast+lossless+KV+compression+for+disaggregated+LLM+serving 7. KVServe: Service-aware KV cache compression for communication-efficient disaggregated LLM serving — Liu, Z., Ma, X., Luo, D., et al., 2026 https://scholar.google.com/scholar?q=KVServe%3A+Service-aware+KV+cache+compression+for+communication-efficient+disaggregated+LLM+serving 8. Efficient lossless compression of scientific floating-point data on CPUs and GPUs (LC / FZ-style pipelines) — Azami, N., Fallin, A., Burtscher, M., 2025 https://scholar.google.com/scholar?q=Efficient+lossless+compression+of+scientific+floating-point+data+on+CPUs+and+GPUs+%28LC+%2F+FZ-style+pipelines%29 9. NVIDIA nvCOMP (ANS, Bitcomp, GDeflate, LZ4) documentation and benchmarks — NVIDIA, 2023-2026 https://scholar.google.com/scholar?q=NVIDIA+nvCOMP+%28ANS%2C+Bitcomp%2C+GDeflate%2C+LZ4%29+documentation+and+benchmarks 10. Lossless compression of BF16 weights: DFloat11 and ZipNN — Zhang et al. (DFloat11); Hershcovich et al. (ZipNN), 2025 https://scholar.google.com/scholar?q=Lossless+compression+of+BF16+weights%3A+DFloat11+and+ZipNN 11. Mooncake: Trading more storage for less computation — Qin, R., et al., 2025 https://scholar.google.com/scholar?q=Mooncake%3A+Trading+more+storage+for+less+computation 12. KVQuant and KIVI — Hooper et al. 2024; Liu et al. 2024, 2024 https://scholar.google.com/scholar?q=KVQuant+and+KIVI Interactive Visualization: FastKVComp: Checking Whether High-Throughput KV Compression Pays Off

   Transcription

6. Cascade: SLO-Aware Latency Budgets for Fair, High-Goodput LLM Serving

   6 novembre 2026À suivre

   This episode examines Cascade, a serving system that uses a single "latency budget" (a request's time-to-first-token target minus its predicted service time) to drive both queue ordering and KV cache …

   Notes de l’épisode

   This episode examines Cascade, a serving system that uses a single "latency budget" (a request's time-to-first-token target minus its predicted service time) to drive both queue ordering and KV cache restoration decisions. It explains why first-come-first-served, shortest-job-first and earliest-deadline-first each fail. The failures are head-of-line blocking, starved long prompts, and an inability to tell a 50 ms job from a 20 s job with the same deadline. The discussion covers goodput, tiered prefix caching across GPU memory, CPU DRAM and NVMe, and chunked prefill. It also covers the paper's motivating findings: aggregate SLO attainment hides starved classes, deep-tier restores such as a 64,000-token NVMe fetch can blow the latency target, and most violations come from queueing rather than a lack of budget. Listeners get a clear walkthrough of the scheduling vocabulary, Jain's fairness index and the budget estimator. They also get a caveat that the headline cluster results come from an extended simulator fed with GB200 NVL72 data-movement profiles rather than a real deployment. Sources: 1. Cascade: Exploiting SLO-Aware latency budget for fair and high goodput LLM inference serving — Muhammad Adnan, Rohan Mahapatra, Prashant J. Nair, Daniel Berger, Pantea Zardoshti, Rodrigo Fonseca, Esha Choukse, 2026 http://arxiv.org/abs/2608.06557 2. Orca: A Distributed Serving System for Transformer-Based Generative Models — Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, Byung-Gon Chun, 2022 https://scholar.google.com/scholar?q=Orca%3A+A+Distributed+Serving+System+for+Transformer-Based+Generative+Models 3. Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve — Amey Agrawal, Nitin Kedia, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav Gulavani, Alexey Tumanov, Ramachandran Ramjee, 2024 https://scholar.google.com/scholar?q=Taming+Throughput-Latency+Tradeoff+in+LLM+Inference+with+Sarathi-Serve 4. Fast Distributed Inference Serving for Large Language Models (FastServe) — Bingyang Wu, Yinmin Zhong, Zili Zhang, Shengyu Liu, Fangyue Liu, Yuanhang Sun, Gang Huang, Xuanzhe Liu, Xin Jin, 2023 https://scholar.google.com/scholar?q=Fast+Distributed+Inference+Serving+for+Large+Language+Models+%28FastServe%29 5. Efficient LLM Scheduling by Learning to Rank — Yichao Fu, Siqi Zhu, Runlong Su, Aurick Qiao, Ion Stoica, Hao Zhang, 2024 https://scholar.google.com/scholar?q=Efficient+LLM+Scheduling+by+Learning+to+Rank 6. DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving — Yinmin Zhong, Shengyu Liu, Junda Chen, Jianbo Hu, Yibo Zhu, Xuanzhe Liu, Xin Jin, Hao Zhang, 2024 https://scholar.google.com/scholar?q=DistServe%3A+Disaggregating+Prefill+and+Decoding+for+Goodput-optimized+Large+Language+Model+Serving 7. Fairness in Serving Large Language Models — Ying Sheng, Shiyi Cao, Dacheng Li, Banghua Zhu, Zhuohan Li, Danyang Zhuo, Joseph E. Gonzalez, Ion Stoica, 2024 https://scholar.google.com/scholar?q=Fairness+in+Serving+Large+Language+Models 8. A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Computer Systems — Raj Jain, Dah-Ming Chiu, William R. Hawe, 1984 https://scholar.google.com/scholar?q=A+Quantitative+Measure+of+Fairness+and+Discrimination+for+Resource+Allocation+in+Shared+Computer+Systems 9. Andes: Defining and Enhancing Quality-of-Experience in LLM-Based Text Streaming Services — Jiachen Liu, Zhiyu Wu, Jae-Won Chung, Fan Lai, Myungjin Lee, Mosharaf Chowdhury, 2024 https://scholar.google.com/scholar?q=Andes%3A+Defining+and+Enhancing+Quality-of-Experience+in+LLM-Based+Text+Streaming+Services 10. Sarathi-Serve: Taming Throughput-Latency Tradeoff in LLM Inference — Agrawal, Kedia, Panwar, Mohan, Kwatra, Gulavani, Tumanov, Ramjee, 2024 https://scholar.google.com/scholar?q=Sarathi-Serve%3A+Taming+Throughput-Latency+Tradeoff+in+LLM+Inference 11. Fairness in Serving Large Language Models (Virtual Token Counter, VTC) — Sheng, Cao, Li, Zhuang, Zhang, Gonzalez, Stoica et al., 2024 https://scholar.google.com/scholar?q=Fairness+in+Serving+Large+Language+Models+%28Virtual+Token+Counter%2C+VTC%29 12. QoServe: Breaking the Silos of LLM Inference Serving — Goel, Mohan, Kwatra, Anupindi, Ramjee, 2026 https://scholar.google.com/scholar?q=QoServe%3A+Breaking+the+Silos+of+LLM+Inference+Serving 13. SLOs-Serve: Optimized Serving of Multi-SLO LLMs — Chen, Jia, Khan, Krishnamurthy, Gibbons, 2025 https://scholar.google.com/scholar?q=SLOs-Serve%3A+Optimized+Serving+of+Multi-SLO+LLMs 14. JITServe: SLO-aware LLM Serving with Imprecise Request Information — Zhang, Wu, Mu, Ning, Liu, Sarda, Lee, Lai, 2025 https://scholar.google.com/scholar?q=JITServe%3A+SLO-aware+LLM+Serving+with+Imprecise+Request+Information 15. FastServe: Iteration-Level Preemptive Scheduling for LLM Inference (MLFQ) — Wu, Zhong, Zhang, Liu, Liu, Sun, Huang, Liu, Jin, 2026 (NSDI) https://scholar.google.com/scholar?q=FastServe%3A+Iteration-Level+Preemptive+Scheduling+for+LLM+Inference+%28MLFQ%29 16. Mooncake: Trading More Storage for Less Computation — Qin, Li, He, Cui, Ren, Zhang, Wu, Zheng, Xu, 2025 https://scholar.google.com/scholar?q=Mooncake%3A+Trading+More+Storage+for+Less+Computation 17. Tutti: Making SSD-backed KV Cache Practical for Long-Context LLM Serving — Qiu, Hu, Wang, Zhu, Yan, Chen, Xu, Chen, Zhang, 2026 https://scholar.google.com/scholar?q=Tutti%3A+Making+SSD-backed+KV+Cache+Practical+for+Long-Context+LLM+Serving 18. LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference — Liu, Cheng, Yao, An, Chen, Feng, Huang, Shen, Zhang, Du, Jiang, 2025 https://scholar.google.com/scholar?q=LMCache%3A+An+Efficient+KV+Cache+Layer+for+Enterprise-Scale+LLM+Inference 19. Scheduling Algorithms for Multiprogramming in a Hard-Real-Time Environment (and least-laxity-first literature, e.g. Mok 1983) — Liu and Layland (1973); Mok, 1973 / 1983 https://scholar.google.com/scholar?q=Scheduling+Algorithms+for+Multiprogramming+in+a+Hard-Real-Time+Environment+%28and+least-laxity-first+literature%2C+e.g.+Mok+1983%29 20. DistServe: Disaggregating Prefill and Decoding for Goodput-Optimized LLM Serving — Zhong, Liu, Chen, Hu, Zhu, Liu, Jin, Zhang, 2024 https://scholar.google.com/scholar?q=DistServe%3A+Disaggregating+Prefill+and+Decoding+for+Goodput-Optimized+LLM+Serving Interactive Visualization: Cascade: SLO-Aware Latency Budgets for Fair, High-Goodput LLM Serving

   Transcription

7. Adaptive Mixed-Precision Computing Speeds Up Nearest Neighbor Search

   5 novembre 2026À suivre

   This episode examines ANNS-AMP, a June 2026 paper proposing adaptive mixed-precision computing for approximate nearest neighbor search. It starts with the IVF-PQ index pipeline: cluster locating, residual …

   Notes de l’épisode

   This episode examines ANNS-AMP, a June 2026 paper proposing adaptive mixed-precision computing for approximate nearest neighbor search. It starts with the IVF-PQ index pipeline: cluster locating, residual calculation, lookup-table construction, distance summation and top-k selection. The paper reports that over 99.9% of distance arithmetic in the cluster-locating and table-construction stages goes to vectors that are not neighbors. The proposed fix is a small runtime predictor that picks the bit-width for each cluster and each query, paired with a bit-serial, bit-plane near-memory accelerator so that fewer bits mean less compute and less memory traffic. The discussion also covers why existing index tuning and early-termination methods miss this redundancy. It questions whether flexible-precision hardware can succeed now that fixed INT8, FP8 and FP4 formats have taken over in neural networks. A running thread asks how much of the reported speed and energy gain comes from the precision idea and how much from the custom silicon. Sources: 1. ANNS-AMP: Accelerating Approximate Nearest Neighbor Search via Adaptive Mixed-Precision Computing — Mingkai Chen, Cheng Liu, Shengwen Liang, Lei Zhang, Xiaowei Li, Huawei Li, 2026 http://arxiv.org/abs/2606.07156 2. Mixed Precision Training — Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, et al., 2018 https://scholar.google.com/scholar?q=Mixed+Precision+Training 3. HAQ: Hardware-Aware Automated Quantization with Mixed Precision — Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, Song Han, 2019 https://scholar.google.com/scholar?q=HAQ%3A+Hardware-Aware+Automated+Quantization+with+Mixed+Precision 4. HAWQ: Hessian Aware Quantization of Neural Networks with Mixed-Precision — Zhen Dong, Zhewei Yao, Amir Gholami, Michael Mahoney, Kurt Keutzer, 2019 https://scholar.google.com/scholar?q=HAWQ%3A+Hessian+Aware+Quantization+of+Neural+Networks+with+Mixed-Precision 5. Stripes: Bit-Serial Deep Neural Network Computing — Patrick Judd, Jorge Albericio, Tayler Hetherington, Tor Aamodt, Andreas Moshovos, 2016 https://scholar.google.com/scholar?q=Stripes%3A+Bit-Serial+Deep+Neural+Network+Computing 6. Bit Fusion: Bit-Level Dynamically Composable Architecture for Accelerating Deep Neural Networks — Hardik Sharma, Jongse Park, Naveen Suda, Liangzhen Lai, Benson Chau, Vikas Chandra, Hadi Esmaeilzadeh, 2018 https://scholar.google.com/scholar?q=Bit+Fusion%3A+Bit-Level+Dynamically+Composable+Architecture+for+Accelerating+Deep+Neural+Networks 7. BISMO: A Scalable Bit-Serial Matrix Multiplication Overlay for Reconfigurable Computing — Yaman Umuroglu, Lahiru Rasnayake, Magnus Sjalander, 2018 https://scholar.google.com/scholar?q=BISMO%3A+A+Scalable+Bit-Serial+Matrix+Multiplication+Overlay+for+Reconfigurable+Computing 8. Product Quantization for Nearest Neighbor Search — Hervé Jégou, Matthijs Douze, Cordelia Schmid, 2011 https://scholar.google.com/scholar?q=Product+Quantization+for+Nearest+Neighbor+Search 9. Billion-scale similarity search with GPUs — Jeff Johnson, Matthijs Douze, Hervé Jégou, 2017 (IEEE Transactions on Big Data, 2019) https://scholar.google.com/scholar?q=Billion-scale+similarity+search+with+GPUs 10. Accelerating Large-Scale Inference with Anisotropic Vector Quantization — Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, Sanjiv Kumar, 2020 https://scholar.google.com/scholar?q=Accelerating+Large-Scale+Inference+with+Anisotropic+Vector+Quantization 11. Co-design Hardware and Algorithm for Vector Search (FANNS) — Wenqi Jiang, Shigang Li, Yu Zhu, Johannes de Fine Licht, Zhenhao He, Runbin Shi, Cédric Renggli, Shuai Zhang, Theodoros Rekatsas, Torsten Hoefler, Gustavo Alonso, 2023 https://scholar.google.com/scholar?q=Co-design+Hardware+and+Algorithm+for+Vector+Search+%28FANNS%29 12. ANSMET: Approximate Nearest Neighbor Search with Near-Memory Processing and Hybrid Early Termination — Yiwei Li, Yuxin Jin, Boyu Tian, Huanchen Zhang, Mingyu Gao, 2025 https://scholar.google.com/scholar?q=ANSMET%3A+Approximate+Nearest+Neighbor+Search+with+Near-Memory+Processing+and+Hybrid+Early+Termination 13. ANNA: Specialized Architecture for Approximate Nearest Neighbor Search — Yejin Lee, Hyunji Choi, Sunhong Min, Hyunseung Lee, Sangwon Beak, Dawoon Jeong, Jae W. Lee, Tae Jun Ham, 2022 https://scholar.google.com/scholar?q=ANNA%3A+Specialized+Architecture+for+Approximate+Nearest+Neighbor+Search 14. JUNO: Optimizing High-Dimensional Approximate Nearest Neighbour Search with Sparsity-Aware Algorithm and Ray-Tracing Core Mapping — Zihan Liu, Wentao Ni, Jingwen Leng, Yu Feng, Cong Guo, Quan Chen, Chao Li, Minyi Guo, Yuhao Zhu, 2024 https://scholar.google.com/scholar?q=JUNO%3A+Optimizing+High-Dimensional+Approximate+Nearest+Neighbour+Search+with+Sparsity-Aware+Algorithm+and+Ray-Tracing+Core+Mapping 15. RaBitQ: Quantizing High-Dimensional Vectors with a Theoretical Error Bound for Approximate Nearest Neighbor Search — Jianyang Gao, Cheng Long, 2024 https://scholar.google.com/scholar?q=RaBitQ%3A+Quantizing+High-Dimensional+Vectors+with+a+Theoretical+Error+Bound+for+Approximate+Nearest+Neighbor+Search 16. High-Dimensional Approximate Nearest Neighbor Search: with Reliable and Efficient Distance Comparison Operations (ADSampling) — Jianyang Gao, Cheng Long, 2023 https://scholar.google.com/scholar?q=High-Dimensional+Approximate+Nearest+Neighbor+Search%3A+with+Reliable+and+Efficient+Distance+Comparison+Operations+%28ADSampling%29 17. Quicker ADC: Unlocking the Hidden Potential of Product Quantization with SIMD (and Fast Scan variants in Faiss) — Fabien André, Anne-Marie Kermarrec, Nicolas Le Scouarnec, 2019 https://scholar.google.com/scholar?q=Quicker+ADC%3A+Unlocking+the+Hidden+Potential+of+Product+Quantization+with+SIMD+%28and+Fast+Scan+variants+in+Faiss%29 Interactive Visualization: Adaptive Mixed-Precision Computing Speeds Up Nearest Neighbor Search

   Transcription

8. Genuine 1-Bit Post-Training Quantization for LLMs: All for 1-Bit

   5 novembre 2026À suivre

   This episode examines "All for 1-Bit," a paper claiming to compress LLM linear weights to exactly one bit each, metadata included, without retraining. It builds on double binary factorization and challenges …

   Notes de l’épisode

   This episode examines "All for 1-Bit," a paper claiming to compress LLM linear weights to exactly one bit each, metadata included, without retraining. It builds on double binary factorization and challenges earlier "1-bit" post-training methods, which land at roughly 2.5 to 3.5 effective bits per weight once scales, masks and outlier indices are counted. The discussion covers why bandwidth-bound decoding motivates extreme quantization, how a product of two binary matrices adds capacity without adding bits, and how null-space reasoning and Shapley-value layer allocation are meant to spend a fixed rank budget well. Along the way it separates post-training from quantization-aware training and sets up the key skeptical test: whether "genuine" and "usable" both hold, given that a score near chance (about 25 on the zero-shot scale) means a broken model. Listeners get a clear explanation of what "bits per weight" really measures and a framework for judging low-bit quantization claims. Sources: 1. All for 1-Bit: Towards Genuine 1-Bit Post-Training Quantization for LLMs — Zhixiong Zhao, Zukang Xu, Guangyu Sun, Lifeng Liu, Dawei Yang, 2026 http://arxiv.org/abs/2609.06161 2. BiLLM: Pushing the Limit of Post-Training Quantization for LLMs — Wei Huang, Yangdong Liu, Haotong Qin, Ying Li, Shiming Zhang, Xianglong Liu, Michele Magno, Xiaojuan Qi, 2024 https://scholar.google.com/scholar?q=BiLLM%3A+Pushing+the+Limit+of+Post-Training+Quantization+for+LLMs 3. ARB-LLM: Alternating Refined Binarizations for Large Language Models — Zhiteng Li, Xianglong Yan, Tianao Zhang, Haotong Qin, Dong Xie, Jiang Tian, Zhongchao Shi, Linghe Kong, Yulun Zhang, Xiaokang Yang, 2024 https://scholar.google.com/scholar?q=ARB-LLM%3A+Alternating+Refined+Binarizations+for+Large+Language+Models 4. OneBit: Towards Extremely Low-bit Large Language Models — Yuzhuang Xu, Xu Han, Zonghan Yang, Shuo Wang, Qingfu Zhu, Zhiyuan Liu, Weidong Liu, Wanxiang Che, 2024 https://scholar.google.com/scholar?q=OneBit%3A+Towards+Extremely+Low-bit+Large+Language+Models 5. BitNet: Scaling 1-bit Transformers for Large Language Models / The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits — Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Huaijie Wang, Lingxiao Ma, Fan Yang, Ruiping Wang, Yi Wu, Furu Wei (BitNet); Shuming Ma et al. (b1.58), 2023 / 2024 https://scholar.google.com/scholar?q=BitNet%3A+Scaling+1-bit+Transformers+for+Large+Language+Models+%2F+The+Era+of+1-bit+LLMs%3A+All+Large+Language+Models+are+in+1.58+Bits 6. GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers — Elias Frantar, Saleh Ashkboos, Torsten Hoefler, Dan Alistarh, 2022 https://scholar.google.com/scholar?q=GPTQ%3A+Accurate+Post-Training+Quantization+for+Generative+Pre-trained+Transformers 7. SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression — Xin Wang, Yu Zheng, Zhongwei Wan, Mi Zhang, 2024 https://scholar.google.com/scholar?q=SVD-LLM%3A+Truncation-aware+Singular+Value+Decomposition+for+Large+Language+Model+Compression 8. AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models — Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, Tat-Seng Chua, 2024 https://scholar.google.com/scholar?q=AlphaEdit%3A+Null-Space+Constrained+Knowledge+Editing+for+Language+Models 9. Training Networks in Null Space of Feature Covariance for Continual Learning — Shipeng Wang, Xiaorong Li, Jian Sun, Zongben Xu, 2021 https://scholar.google.com/scholar?q=Training+Networks+in+Null+Space+of+Feature+Covariance+for+Continual+Learning 10. A Unified Approach to Interpreting Model Predictions (SHAP) — Scott M. Lundberg, Su-In Lee, 2017 https://scholar.google.com/scholar?q=A+Unified+Approach+to+Interpreting+Model+Predictions+%28SHAP%29 11. Data Shapley: Equitable Valuation of Data for Machine Learning — Amirata Ghorbani, James Zou, 2019 https://scholar.google.com/scholar?q=Data+Shapley%3A+Equitable+Valuation+of+Data+for+Machine+Learning 12. HAWQ-V2: Hessian Aware trace-Weighted Quantization of Neural Networks — Zhen Dong, Zhewei Yao, Yaohui Cai, Daiyaan Arfeen, Amir Gholami, Michael W. Mahoney, Kurt Keutzer, 2020 https://scholar.google.com/scholar?q=HAWQ-V2%3A+Hessian+Aware+trace-Weighted+Quantization+of+Neural+Networks 13. SqueezeLLM: Dense-and-Sparse Quantization — Sehoon Kim, Coleman Hooper, Amir Gholami, Zhen Dong, Xiuyu Li, Sheng Shen, Michael W. Mahoney, Kurt Keutzer, 2023 https://scholar.google.com/scholar?q=SqueezeLLM%3A+Dense-and-Sparse+Quantization 14. Addition is almost all you need: Compressing neural networks with double binary factorization — Vladimír Boža, Vladimír Macko, 2025 https://scholar.google.com/scholar?q=Addition+is+almost+all+you+need%3A+Compressing+neural+networks+with+double+binary+factorization 15. LittleBit: Ultra Low-Bit Quantization via Latent Factorization — Lee et al., 2026 https://scholar.google.com/scholar?q=LittleBit%3A+Ultra+Low-Bit+Quantization+via+Latent+Factorization 16. The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits (BitNet b1.58) — Shuming Ma et al., 2024 https://scholar.google.com/scholar?q=The+Era+of+1-bit+LLMs%3A+All+Large+Language+Models+are+in+1.58+Bits+%28BitNet+b1.58%29 17. FBI-LLM: Scaling Up Fully Binarized LLMs from Scratch via Autoregressive Distillation — Chuang Ma et al., 2024 https://scholar.google.com/scholar?q=FBI-LLM%3A+Scaling+Up+Fully+Binarized+LLMs+from+Scratch+via+Autoregressive+Distillation 18. HBLLM: Haar-based approach for accurate structured 1-bit quantized LLMs — Chen et al., 2025 https://scholar.google.com/scholar?q=HBLLM%3A+Haar-based+approach+for+accurate+structured+1-bit+quantized+LLMs 19. AQLM: Extreme Compression of LLMs via Additive Quantization — Egiazarian et al., 2024 https://scholar.google.com/scholar?q=AQLM%3A+Extreme+Compression+of+LLMs+via+Additive+Quantization 20. QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks — Tseng et al., 2024 https://scholar.google.com/scholar?q=QuIP%23%3A+Even+Better+LLM+Quantization+with+Hadamard+Incoherence+and+Lattice+Codebooks 21. QTIP: Quantization with Trellises and Incoherence Processing — Tseng et al., 2024 https://scholar.google.com/scholar?q=QTIP%3A+Quantization+with+Trellises+and+Incoherence+Processing 22. Extreme Compression of Large Language Models via Additive Quantization / PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression — Malinovskii et al., 2024 https://scholar.google.com/scholar?q=Extreme+Compression+of+Large+Language+Models+via+Additive+Quantization+%2F+PV-Tuning%3A+Beyond+Straight-Through+Estimation+for+Extreme+LLM+Compression 23. SVD-LLM: Truncation-aware Singular Value Decomposition for LLM Compression — Xin Wang et al., 2024 https://scholar.google.com/scholar?q=SVD-LLM%3A+Truncation-aware+Singular+Value+Decomposition+for+LLM+Compression 24. BitNet a4.8 / bitnet.cpp: Efficient edge inference for ternary LLMs — Microsoft Research, 2024-2025 https://scholar.google.com/scholar?q=BitNet+a4.8+%2F+bitnet.cpp%3A+Efficient+edge+inference+for+ternary+LLMs 25. ParetoQ: Scaling Laws in Extremely Low-bit LLM Quantization — Zechun Liu et al., 2025 https://scholar.google.com/scholar?q=ParetoQ%3A+Scaling+Laws+in+Extremely+Low-bit+LLM+Quantization Interactive Visualization: Genuine 1-Bit Post-Training Quantization for LLMs: All for 1-Bit

   Transcription

9. PagedMoE: Cross-Precision Expert Caching for SSD-Resident MoE Inference

   5 novembre 2026À suivre

   This episode examines PagedMoE, a system for running a huge mixture-of-experts model (Qwen3.5-397B-A17B) on a single PC with expert weights stored on an SSD. Cache misses leave decoding stalled on storage …

   Notes de l’épisode

   This episode examines PagedMoE, a system for running a huge mixture-of-experts model (Qwen3.5-397B-A17B) on a single PC with expert weights stored on an SSD. Cache misses leave decoding stalled on storage nearly half the time, and giving each token a different bit-width per expert would normally make that worse. Treating every expert-and-precision pair as its own cache object drops the hit rate from 57% to 35%. The paper's fix is to store each expert as nested refinement units, so a resident low-precision copy can be topped up with only the missing bits, or reused with no I/O at all. The discussion separates this idea from the prior work it borrows from (Matryoshka Quantization, Any-Precision LLM, SliceMoE, HOBBIT). It also splits the reported 1.56x speedup into the part from requesting fewer bits and the part from sharing cached bits across precisions. Listeners get a clear walk through the binary-tree codebook construction, the quantization-error numbers, and the limits of the evidence, such as a single-layer headroom study and no head-to-head comparison with SliceMoE. Sources: 1. PagedMoE: Cross-Precision Expert Caching for Token-Level Dynamic Precision in SSD-Resident MoE Inference — Yutong Lin, Yingdi Shan, Shaoyuan Chen, Yuanyong Chen, Yongwei Wu, Mingxing Zhang, 2026 http://arxiv.org/abs/2611.03173 2. Fast Inference of Mixture-of-Experts Language Models with Offloading — Artyom Eliseev, Denis Mazur, 2023 https://scholar.google.com/scholar?q=Fast+Inference+of+Mixture-of-Experts+Language+Models+with+Offloading 3. MoE-Infinity: Efficient MoE Inference on Personal Machines with Sparsity-Aware Expert Cache — Leyang Xue, Yao Fu, Zhan Lu, Luo Mai, Mahesh Marina, 2024 (arXiv; later revisions) https://scholar.google.com/scholar?q=MoE-Infinity%3A+Efficient+MoE+Inference+on+Personal+Machines+with+Sparsity-Aware+Expert+Cache 4. EdgeMoE: Empowering Sparse Large Language Models on Mobile Devices — Rongjie Yi, Liwei Guo, Shiyun Wei, Ao Zhou, Shangguang Wang, Mengwei Xu, 2023 https://scholar.google.com/scholar?q=EdgeMoE%3A+Empowering+Sparse+Large+Language+Models+on+Mobile+Devices 5. LLM in a Flash: Efficient Large Language Model Inference with Limited Memory — Keivan Alizadeh, Iman Mirzadeh, et al. (Apple), 2023 https://scholar.google.com/scholar?q=LLM+in+a+Flash%3A+Efficient+Large+Language+Model+Inference+with+Limited+Memory 6. HOBBIT: A Mixed Precision Expert Offloading System for Fast MoE Inference — Peng Tang, Jiacheng Liu, Xiaofeng Hou, et al., 2024 https://scholar.google.com/scholar?q=HOBBIT%3A+A+Mixed+Precision+Expert+Offloading+System+for+Fast+MoE+Inference 7. Mixture Compressor for Mixture-of-Experts LLMs Gains More (MC) — Wei Huang, Yue Liao, Jianhui Liu, et al., 2024 https://scholar.google.com/scholar?q=Mixture+Compressor+for+Mixture-of-Experts+LLMs+Gains+More+%28MC%29 8. GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers — Elias Frantar, Saleh Ashkboos, Torsten Hoefler, Dan Alistarh, 2022 https://scholar.google.com/scholar?q=GPTQ%3A+Accurate+Post-Training+Quantization+for+Generative+Pre-trained+Transformers 9. Microscaling Data Formats for Deep Learning (OCP MX specification) — Bita Darvish Rouhani et al. (Microsoft, AMD, Arm, Intel, Meta, NVIDIA, Qualcomm), 2023 https://scholar.google.com/scholar?q=Microscaling+Data+Formats+for+Deep+Learning+%28OCP+MX+specification%29 10. Any-Precision LLM: Low-Cost Deployment of Multiple, Different-Sized LLMs — Yeonhong Park, Jake Hyun, SangLyul Cho, Bonggeun Sim, Jae W. Lee, 2024 https://scholar.google.com/scholar?q=Any-Precision+LLM%3A+Low-Cost+Deployment+of+Multiple%2C+Different-Sized+LLMs 11. Matryoshka Quantization — Pranav Nair, Puranjay Datta, Jeff Dean, Prateek Jain, Aditya Kusupati, 2025 https://scholar.google.com/scholar?q=Matryoshka+Quantization 12. QTIP: Quantization with Trellises and Incoherence Processing — Albert Tseng, Qingyao Sun, David Hou, Christopher De Sa, 2024 https://scholar.google.com/scholar?q=QTIP%3A+Quantization+with+Trellises+and+Incoherence+Processing 13. Stripes: Bit-Serial Deep Neural Network Computing — Patrick Judd, Jorge Albericio, Tayler Hetherington, Tor Aamodt, Andreas Moshovos, 2016 https://scholar.google.com/scholar?q=Stripes%3A+Bit-Serial+Deep+Neural+Network+Computing 14. SliceMoE: Bit-Sliced Expert Caching under Miss-Rate Constraints for Efficient MoE Inference — Yuseon Choi, Sangjin Kim, Jungjun Oh, Gwangtae Park, Byeongcheol Kim, Hoi-Jun Yoo, 2026 https://scholar.google.com/scholar?q=SliceMoE%3A+Bit-Sliced+Expert+Caching+under+Miss-Rate+Constraints+for+Efficient+MoE+Inference 15. D²MoE: Dual Routing and Dynamic Scheduling for Efficient On-Device MoE-based LLM Serving — Haodong Wang, Qihua Zhou, Zicong Hong, Song Guo, 2025 https://scholar.google.com/scholar?q=D%C2%B2MoE%3A+Dual+Routing+and+Dynamic+Scheduling+for+Efficient+On-Device+MoE-based+LLM+Serving 16. DyMoE: Dynamic Expert Orchestration with Mixed-Precision Quantization for Efficient MoE Inference on Edge — Yuegui Huang, Zhiyuan Fang, Weiqi Luo, Ruoyu Wu, Wuhui Chen, Zibin Zheng, 2026 https://scholar.google.com/scholar?q=DyMoE%3A+Dynamic+Expert+Orchestration+with+Mixed-Precision+Quantization+for+Efficient+MoE+Inference+on+Edge 17. MxMoE: Mixed-precision Quantization for MoE with Accuracy and Performance Co-Design — Haojie Duanmu, Xiuhong Li, Zhihang Yuan, Size Zheng, Jiangfei Duan, Xingcheng Zhang, Dahua Lin, 2025 https://scholar.google.com/scholar?q=MxMoE%3A+Mixed-precision+Quantization+for+MoE+with+Accuracy+and+Performance+Co-Design 18. GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs — Jianing Deng, Song Wang, Dongwei Wang, Zijie Liu, Tianlong Chen, Huanrui Yang, Jingtong Hu, 2026 https://scholar.google.com/scholar?q=GEMQ%3A+Global+Expert-Level+Mixed-Precision+Quantization+for+MoE+LLMs 19. KTransformers: Unleashing the Full Potential of CPU/GPU Hybrid Inference for MoE Models — Hongtao Chen, Weiyu Xie, Boxin Zhang, et al., 2025 https://scholar.google.com/scholar?q=KTransformers%3A+Unleashing+the+Full+Potential+of+CPU%2FGPU+Hybrid+Inference+for+MoE+Models 20. T-MAC: CPU Renaissance via Table Lookup for Low-Bit LLM Deployment on Edge — Jianyu Wei, Shijie Cao, Ting Cao, Lingxiao Ma, Lei Wang, Yanyong Zhang, Mao Yang, 2025 https://scholar.google.com/scholar?q=T-MAC%3A+CPU+Renaissance+via+Table+Lookup+for+Low-Bit+LLM+Deployment+on+Edge 21. FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU — Ying Sheng et al., 2023 https://scholar.google.com/scholar?q=FlexGen%3A+High-Throughput+Generative+Inference+of+Large+Language+Models+with+a+Single+GPU 22. Efficient Memory Management for Large Language Model Serving with PagedAttention — Woosuk Kwon et al., 2023 https://scholar.google.com/scholar?q=Efficient+Memory+Management+for+Large+Language+Model+Serving+with+PagedAttention 23. Speculative MoE / expert prefetching works (e.g., Pre-gated MoE, ProMoE, AdapMoE) — Various, 2024–2025 https://scholar.google.com/scholar?q=Speculative+MoE+%2F+expert+prefetching+works+%28e.g.%2C+Pre-gated+MoE%2C+ProMoE%2C+AdapMoE%29 Interactive Visualization: PagedMoE: Cross-Precision Expert Caching for SSD-Resident MoE Inference

   Transcription

10. LeanSparse: Shrinking GPU Memory for Sparse-Attention LLM Serving

    4 novembre 2026À suivre

    This episode examines LeanSparse, a serving system that cuts the GPU memory needed to hold many long-context conversations on models with native sparse attention, such as DeepSeek-V4-Flash. It explains why …

    Notes de l’épisode

    This episode examines LeanSparse, a serving system that cuts the GPU memory needed to hold many long-context conversations on models with native sparse attention, such as DeepSeek-V4-Flash. It explains why sparse attention reduces reads but not storage. Even with KV offloaded to host memory, the lightning indexer's keys must stay on the GPU, and at 512K tokens they take about 7.2 times the space of the hot KV buffer. It also covers why agent-style workloads, where context grows every turn, make this memory limit bite, and why prefill needs a large temporary workspace. That workspace forces a bad choice between reserving memory that sits idle and making new requests queue behind decoders, which hurts time to first token. The proposed fixes are to stream each layer's history through two alternating GPU buffers during prefill, and to search compact indexer keys during decode, fetching only a short list of originals to rank exactly. All of this runs on a shared instance, avoiding the duplicated weights of prefill-decode disaggregation. Listeners get a clear walkthrough of where memory goes in modern sparse-attention serving, plus the reasoning behind why the obvious fix of offloading the indexer keys fails. Sources: 1. LeanSparse: Memory-Efficient Serving for Native Sparse-Attention LLMs — Zaifeng Pan, Chris Wu, Xinwei Qiang, Zhengding Hu, Zhongkai Yu, Yufei Ding, 2026 http://arxiv.org/abs/2611.04671 2. Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention — Jingyang Yuan, Huazuo Gao, Damai Dai, et al. (DeepSeek-AI, Peking University), 2025 https://scholar.google.com/scholar?q=Native+Sparse+Attention%3A+Hardware-Aligned+and+Natively+Trainable+Sparse+Attention 3. DeepSeek-V3.2-Exp technical report (DeepSeek Sparse Attention, DSA) — DeepSeek-AI, 2025 https://scholar.google.com/scholar?q=DeepSeek-V3.2-Exp+technical+report+%28DeepSeek+Sparse+Attention%2C+DSA%29 4. MoBA: Mixture of Block Attention for Long-Context LLMs — Enzhe Lu, Zhejun Jiang, Jingyuan Liu, et al. (Moonshot AI), 2025 https://scholar.google.com/scholar?q=MoBA%3A+Mixture+of+Block+Attention+for+Long-Context+LLMs 5. Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference — Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, Song Han, 2024 https://scholar.google.com/scholar?q=Quest%3A+Query-Aware+Sparsity+for+Efficient+Long-Context+LLM+Inference 6. InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management — Wonbeom Lee, Jungi Lee, Junghwan Seo, Jaewoong Sim, 2024 https://scholar.google.com/scholar?q=InfiniGen%3A+Efficient+Generative+Inference+of+Large+Language+Models+with+Dynamic+KV+Cache+Management 7. ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference — Hanshi Sun, Li-Wen Chang, Wenlei Bao, et al., 2024 https://scholar.google.com/scholar?q=ShadowKV%3A+KV+Cache+in+Shadows+for+High-Throughput+Long-Context+LLM+Inference 8. RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval — Di Liu, Meng Chen, Baotong Lu, et al., 2024 https://scholar.google.com/scholar?q=RetrievalAttention%3A+Accelerating+Long-Context+LLM+Inference+via+Vector+Retrieval 9. Billion-scale similarity search with GPUs (FAISS) — Jeff Johnson, Matthijs Douze, Hervé Jégou, 2017 https://scholar.google.com/scholar?q=Billion-scale+similarity+search+with+GPUs+%28FAISS%29 10. Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving — Ruoyu Qin, Zheming Li, Weiran He, et al. (Moonshot AI, Tsinghua), 2024 https://scholar.google.com/scholar?q=Mooncake%3A+A+KVCache-centric+Disaggregated+Architecture+for+LLM+Serving 11. DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving — Yinmin Zhong, Shengyu Liu, Junda Chen, et al., 2024 https://scholar.google.com/scholar?q=DistServe%3A+Disaggregating+Prefill+and+Decoding+for+Goodput-optimized+Large+Language+Model+Serving 12. Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve — Amey Agrawal, Nitin Kedia, Ashish Panwar, et al., 2024 https://scholar.google.com/scholar?q=Taming+Throughput-Latency+Tradeoff+in+LLM+Inference+with+Sarathi-Serve 13. FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU — Ying Sheng, Lianmin Zheng, Binhang Yuan, et al., 2023 https://scholar.google.com/scholar?q=FlexGen%3A+High-Throughput+Generative+Inference+of+Large+Language+Models+with+a+Single+GPU 14. PQCache: Product Quantization-based KVCache for Long Context LLM Inference — Hailin Zhang et al., 2025 https://scholar.google.com/scholar?q=PQCache%3A+Product+Quantization-based+KVCache+for+Long+Context+LLM+Inference 15. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models (DeepSeek Sparse Attention) — DeepSeek-AI, 2025 https://scholar.google.com/scholar?q=DeepSeek-V3.2%3A+Pushing+the+Frontier+of+Open+Large+Language+Models+%28DeepSeek+Sparse+Attention%29 16. CachedAttention (AttentionStore): Cost-Efficient LLM Serving for Multi-Turn Conversations — Bin Gao et al., 2024 https://scholar.google.com/scholar?q=CachedAttention+%28AttentionStore%29%3A+Cost-Efficient+LLM+Serving+for+Multi-Turn+Conversations 17. SparseServe: Unlocking Parallelism for Dynamic Sparse Attention in Long-Context LLM Serving — Zhou et al., 2025 https://scholar.google.com/scholar?q=SparseServe%3A+Unlocking+Parallelism+for+Dynamic+Sparse+Attention+in+Long-Context+LLM+Serving 18. Sarathi-Serve: Taming Throughput-Latency Tradeoff in LLM Inference — Agrawal et al., 2024 https://scholar.google.com/scholar?q=Sarathi-Serve%3A+Taming+Throughput-Latency+Tradeoff+in+LLM+Inference 19. Product Quantization for Nearest Neighbor Search — Hervé Jégou, Matthijs Douze, Cordelia Schmid, 2011 https://scholar.google.com/scholar?q=Product+Quantization+for+Nearest+Neighbor+Search Interactive Visualization: LeanSparse: Shrinking GPU Memory for Sparse-Attention LLM Serving

    Transcription

11. ECHO: Lossless KV Cache Offloading for Sparse Attention LLMs

    4 novembre 2026À suivre

    This episode examines ECHO, a system from Shanghai Jiao Tong University and collaborators (OSDI 2026) that offloads the KV cache to host memory for serving sparse-attention models like DeepSeek-V3.2. It …

    Notes de l’épisode

    This episode examines ECHO, a system from Shanghai Jiao Tong University and collaborators (OSDI 2026) that offloads the KV cache to host memory for serving sparse-attention models like DeepSeek-V3.2. It explains why sparse attention cuts compute but not storage. Cache size still grows linearly with context, so GPU memory limits batch size and leaves the sparse kernels underused. It then covers why naive recall over PCIe is slow: the top-k token set is known only after the lightning indexer has scored every earlier token. Prior prefetchers such as InfiniGen and FreeKV guess which tokens to fetch, which is lossy. ECHO instead guesses only when to fetch and then guarantees a recall of any missed token. To make this work under CUDA graph execution, it keeps its cache-management state in fixed-size integer tensors on the GPU. The discussion also stresses that the predictability ECHO relies on comes from one trained indexer, not from attention in general. It ends by questioning whether "lossless" is only a design property or something the paper actually measures. Sources: 1. ECHO: Lossless KV Cache Offloading for Sparse Attention LLMs https://www.usenix.org/system/files/osdi26-liu-guangda.pdf 2. Efficient Memory Management for Large Language Model Serving with PagedAttention — Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, et al., 2023 https://scholar.google.com/scholar?q=Efficient+Memory+Management+for+Large+Language+Model+Serving+with+PagedAttention 3. InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management — Wonbeom Lee, Jungi Lee, Junghwan Seo, Jaewoong Sim, 2024 https://scholar.google.com/scholar?q=InfiniGen%3A+Efficient+Generative+Inference+of+Large+Language+Models+with+Dynamic+KV+Cache+Management 4. SGLang: Efficient Execution of Structured Language Model Programs — Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, et al., 2024 https://scholar.google.com/scholar?q=SGLang%3A+Efficient+Execution+of+Structured+Language+Model+Programs 5. FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU — Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, et al., 2023 https://scholar.google.com/scholar?q=FlexGen%3A+High-Throughput+Generative+Inference+of+Large+Language+Models+with+a+Single+GPU 6. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness — Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, Christopher Ré, 2022 https://scholar.google.com/scholar?q=FlashAttention%3A+Fast+and+Memory-Efficient+Exact+Attention+with+IO-Awareness 7. FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision — Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao, 2024 https://scholar.google.com/scholar?q=FlashAttention-3%3A+Fast+and+Accurate+Attention+with+Asynchrony+and+Low-precision 8. DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention — DeepSeek-AI, 2025 https://scholar.google.com/scholar?q=DeepSeek-V3.2-Exp%3A+Boosting+Long-Context+Efficiency+with+DeepSeek+Sparse+Attention 9. Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention — Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, et al. (DeepSeek-AI and collaborators), 2025 https://scholar.google.com/scholar?q=Native+Sparse+Attention%3A+Hardware-Aligned+and+Natively+Trainable+Sparse+Attention 10. FreeKV: Boosting KV Cache Retrieval for Efficient LLM Inference — Liu et al., 2025 https://scholar.google.com/scholar?q=FreeKV%3A+Boosting+KV+Cache+Retrieval+for+Efficient+LLM+Inference 11. ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction — Chen et al., 2024 https://scholar.google.com/scholar?q=ArkVale%3A+Efficient+Generative+LLM+Inference+with+Recallable+Key-Value+Eviction 12. SparseServe: Unlocking Parallelism for Dynamic Sparse Attention in Long-Context LLM Serving — Zhou et al., 2025 https://scholar.google.com/scholar?q=SparseServe%3A+Unlocking+Parallelism+for+Dynamic+Sparse+Attention+in+Long-Context+LLM+Serving 13. ESS: An Offload-Centric Latent-Cache Management Architecture for DeepSeek-V3.2-Exp — Chen et al., 2025 https://scholar.google.com/scholar?q=ESS%3A+An+Offload-Centric+Latent-Cache+Management+Architecture+for+DeepSeek-V3.2-Exp 14. Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference — Tang et al., 2024 https://scholar.google.com/scholar?q=Quest%3A+Query-Aware+Sparsity+for+Efficient+Long-Context+LLM+Inference 15. Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention (NSA) — Yuan et al. (DeepSeek-AI), 2025 https://scholar.google.com/scholar?q=Native+Sparse+Attention%3A+Hardware-Aligned+and+Natively+Trainable+Sparse+Attention+%28NSA%29 16. LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference — Cheng et al., 2025 https://scholar.google.com/scholar?q=LMCache%3A+An+Efficient+KV+Cache+Layer+for+Enterprise-Scale+LLM+Inference 17. Sarathi-Serve: Taming Throughput-Latency Tradeoff in LLM Inference — Agrawal et al., 2024 https://scholar.google.com/scholar?q=Sarathi-Serve%3A+Taming+Throughput-Latency+Tradeoff+in+LLM+Inference Interactive Visualization: ECHO: Lossless KV Cache Offloading for Sparse Attention LLMs

    Transcription

12. Joules-to-GDP: Economic Value Produced Per Unit of AI Energy

    4 novembre 2026À suivre

    This episode examines "Joules-to-GDP," a Stanford-led paper that asks how much economic value AI systems produce per joule of energy when doing real professional work. It lays out the setup: 220 public GDPval …

    Notes de l’épisode

    This episode examines "Joules-to-GDP," a Stanford-led paper that asks how much economic value AI systems produce per joule of energy when doing real professional work. It lays out the setup: 220 public GDPval tasks across 44 occupations, expert completion time multiplied by Bureau of Labor Statistics median wages as the dollar value, and a small LLM judge deciding whether the AI's deliverable beats the expert's. The central metrics are Net Value per Joule and Time Saving per Joule. Both are measured against an expert-only baseline and a tougher expert-with-AI baseline. The discussion also covers how energy is metered across a whole batch and split among concurrent tasks, and how running a small model across eight GPUs can waste power. It flags a confound: one model ran on H200s rather than H100s, so differences between models may reflect hardware rather than capability. Listeners get a sceptical look at whether ranking AI by energy efficiency gives a different answer from ranking by capability, cost or speed. Sources: 1. Joules-to-GDP: Economic Value Produced Per Unit of AI Energy https://arxiv.org/pdf/2610.30858 2. GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks — Tejal Patwardhan, Rachel Dias, Elizabeth Proehl, Grace Kim, et al. (OpenAI), 2025 https://scholar.google.com/scholar?q=GDPval%3A+Evaluating+AI+Model+Performance+on+Real-World+Economically+Valuable+Tasks 3. The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization — Jae-Won Chung, Jeff J. Ma, Ruofan Wu, Jiachen Liu, Oh Jun Kweon, Yuxuan Xia, Zhiyu Wu, Mosharaf Chowdhury, 2025 https://scholar.google.com/scholar?q=The+ML.ENERGY+Benchmark%3A+Toward+Automated+Inference+Energy+Measurement+and+Optimization 4. From Words to Watts: Benchmarking the Energy Costs of Large Language Model Inference — Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, Vijay Gadepally, 2023 https://scholar.google.com/scholar?q=From+Words+to+Watts%3A+Benchmarking+the+Energy+Costs+of+Large+Language+Model+Inference 5. Energy and Policy Considerations for Deep Learning in NLP — Emma Strubell, Ananya Ganesh, Andrew McCallum, 2019 https://scholar.google.com/scholar?q=Energy+and+Policy+Considerations+for+Deep+Learning+in+NLP 6. The Simple Macroeconomics of AI — Daron Acemoglu, 2024 https://scholar.google.com/scholar?q=The+Simple+Macroeconomics+of+AI 7. Cost-of-Pass: An Economic Framework for Evaluating Language Models — Mehmet Hamza Erol, Batu El, Mirac Suzgun, Mert Yuksekgonul, James Zou, 2026 https://scholar.google.com/scholar?q=Cost-of-Pass%3A+An+Economic+Framework+for+Evaluating+Language+Models 8. Intelligence per Watt: Measuring Intelligence Efficiency of Local AI — Jon Saad-Falcon, Avanika Narayan, et al., 2025 https://scholar.google.com/scholar?q=Intelligence+per+Watt%3A+Measuring+Intelligence+Efficiency+of+Local+AI 9. MLPerf Power: Benchmarking the Energy Efficiency of Machine Learning Systems from Microwatts to Megawatts — Arya Tschand et al., 2024 https://scholar.google.com/scholar?q=MLPerf+Power%3A+Benchmarking+the+Energy+Efficiency+of+Machine+Learning+Systems+from+Microwatts+to+Megawatts 10. Economic Evaluations of Language Models — Alexander Wan, Stephane Hatgis-Kessell, Tomás Aguirre, Percy Liang, Rishi Bommasani, 2026 https://scholar.google.com/scholar?q=Economic+Evaluations+of+Language+Models 11. Remote Labor Index: Measuring AI Automation of Remote Work — Mantas Mazeika et al., 2025 https://scholar.google.com/scholar?q=Remote+Labor+Index%3A+Measuring+AI+Automation+of+Remote+Work 12. Power Hungry Processing: Watts Driving the Cost of AI Deployment? — Alexandra Sasha Luccioni, Yacine Jernite, Emma Strubell, 2024 https://scholar.google.com/scholar?q=Power+Hungry+Processing%3A+Watts+Driving+the+Cost+of+AI+Deployment%3F 13. AI Agents That Matter — Sayash Kapoor et al., 2024 https://scholar.google.com/scholar?q=AI+Agents+That+Matter 14. Generative AI at Work — Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond, 2025 https://scholar.google.com/scholar?q=Generative+AI+at+Work 15. Characterization of Request and Token Energy Costs for LLM Inference Workloads on GPU Platforms — Prabhu Vellaisamy et al., 2026 https://scholar.google.com/scholar?q=Characterization+of+Request+and+Token+Energy+Costs+for+LLM+Inference+Workloads+on+GPU+Platforms Interactive Visualization: Joules-to-GDP: Economic Value Produced Per Unit of AI Energy

    Transcription

13. Intelligence per Watt: Measuring Local AI Efficiency Against the Cloud

    4 novembre 2026À suivre

    This episode examines "Intelligence per Watt," a Stanford and Together AI paper that asks whether small local models on laptop-class hardware can take real query load off the cloud. It walks through the paper's …

    Notes de l’épisode

    This episode examines "Intelligence per Watt," a Stanford and Together AI paper that asks whether small local models on laptop-class hardware can take real query load off the cloud. It walks through the paper's definitions. "Local" means at most 20 billion active parameters per token, which includes mixture-of-experts models like gpt-oss-120B. The proposed metric is average task accuracy divided by average power draw, and the paper also reports an energy-per-query (joules) variant. The discussion points out that watts alone ignore latency, so a low-power device that runs slowly can still cost more energy per query. It also flags the measurement caveats: power is read from the accelerator only, not the wall; queries run at batch size 1; chat accuracy is judged against a single reference model's answers; and "best-of-local" coverage is an oracle router that couldn't be deployed. The study spans about a million single-turn queries, more than twenty models and eight accelerators, and the episode places it against earlier LLM energy work such as "From Words to Watts" and MLPerf Power. Listeners get a careful guide to what the metric does and doesn't capture before the headline efficiency numbers are discussed. Sources: 1. Intelligence per Watt: Measuring Intelligence Efficiency of Local AI — Jon Saad-Falcon, Avanika Narayan, Hakki Orhun Akengin, J. Wes Griffin, Herumb Shandilya, Adrian Gamarra Lafuente, Medhya Goel, Rebecca Joseph, Shlok Natarajan, Etash Kumar Guha, Shang Zhu, Ben Athiwaratkun, John Hennessy, Azalia Mirhoseini, Christopher Ré, 2025 http://arxiv.org/abs/2511.07885 2. Intelligence per Watt: Measuring Intelligence Efficiency of Local AI — Jon Saad-Falcon, Avanika Narayan, Hakki Orhun Akengin, et al. (Stanford, Together AI; senior authors incl. John Hennessy, Azalia Mirhoseini, Christopher Ré), 2025 (v7 revised Oct 2026) https://scholar.google.com/scholar?q=Intelligence+per+Watt%3A+Measuring+Intelligence+Efficiency+of+Local+AI 3. MLPerf Power: Benchmarking the Energy Efficiency of Machine Learning Systems from Microwatts to Megawatts for Sustainable AI — Arya Tschand, Vijay Janapa Reddi, et al. (MLCommons), 2024 https://scholar.google.com/scholar?q=MLPerf+Power%3A+Benchmarking+the+Energy+Efficiency+of+Machine+Learning+Systems+from+Microwatts+to+Megawatts+for+Sustainable+AI 4. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model — Alexandra Sasha Luccioni, Sylvain Viguier, Anne-Laure Ligozat, 2022 https://scholar.google.com/scholar?q=Estimating+the+Carbon+Footprint+of+BLOOM%2C+a+176B+Parameter+Language+Model 5. From Words to Watts: Benchmarking the Energy Costs of Large Language Model Inference — Siddharth Samsi, Dan Zhao, Joseph McDonald, et al. (MIT Lincoln Laboratory), 2023 https://scholar.google.com/scholar?q=From+Words+to+Watts%3A+Benchmarking+the+Energy+Costs+of+Large+Language+Model+Inference 6. Small Language Models are the Future of Agentic AI — Peter Belcak, Greg Heinrich, Shizhe Diao, et al. (NVIDIA Research), 2025 https://scholar.google.com/scholar?q=Small+Language+Models+are+the+Future+of+Agentic+AI 7. Efficient Memory Management for Large Language Model Serving with PagedAttention — Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, et al., 2023 https://scholar.google.com/scholar?q=Efficient+Memory+Management+for+Large+Language+Model+Serving+with+PagedAttention 8. AWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration — Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Song Han, et al., 2023 (MLSys 2024) https://scholar.google.com/scholar?q=AWQ%3A+Activation-aware+Weight+Quantization+for+On-Device+LLM+Compression+and+Acceleration 9. Energy Considerations of Large Language Model Inference and Efficiency Optimizations — Jared Fernandez, Clara Na, Vashisth Tiwari, Yonatan Bisk, Sasha Luccioni, Emma Strubell, 2025 https://scholar.google.com/scholar?q=Energy+Considerations+of+Large+Language+Model+Inference+and+Efficiency+Optimizations 10. Part-time Power Measurements: nvidia-smi's Lack of Attention — Zeyu Yang, Karel Adamek, Wesley Armour, 2023 https://scholar.google.com/scholar?q=Part-time+Power+Measurements%3A+nvidia-smi%27s+Lack+of+Attention 11. Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute — Felipe Oviedo et al., 2025 https://scholar.google.com/scholar?q=Energy+Use+of+AI+Inference%3A+Efficiency+Pathways+and+Test-Time+Compute 12. Minions: Cost-Efficient Collaboration Between On-Device and Cloud Language Models — Avanika Narayan, Dan Biderman, Sabri Eyuboglu, Avner May, Scott Linderman, James Zou, Christopher Ré, 2025 https://scholar.google.com/scholar?q=Minions%3A+Cost-Efficient+Collaboration+Between+On-Device+and+Cloud+Language+Models 13. RouteLLM: Learning to Route LLMs with Preference Data — Isaac Ong et al., 2025 https://scholar.google.com/scholar?q=RouteLLM%3A+Learning+to+Route+LLMs+with+Preference+Data 14. PowerInfer: Fast Large Language Model Serving with a Consumer-Grade GPU — Yixin Song, Zeyu Mi, Haotong Xie, Haibo Chen, 2024 https://scholar.google.com/scholar?q=PowerInfer%3A+Fast+Large+Language+Model+Serving+with+a+Consumer-Grade+GPU 15. Green AI — Roy Schwartz, Jesse Dodge, Noah A. Smith, Oren Etzioni, 2020 https://scholar.google.com/scholar?q=Green+AI 16. Implications of Historical Trends in the Electrical Efficiency of Computing — Jonathan Koomey, Stephen Berard, Marla Sanchez, Henry Wong, 2010 https://scholar.google.com/scholar?q=Implications+of+Historical+Trends+in+the+Electrical+Efficiency+of+Computing Interactive Visualization: Intelligence per Watt: Measuring Local AI Efficiency Against the Cloud

    Transcription

14. How to Loop Your Transformer: Scaling, Specialization, Parallel Inference

    3 novembre 2026À suivre

    This episode examines how to loop a transformer's layers, meaning reuse the same weights several times per token, depending on whether model size, training compute, or decoding speed is the scarce resource. It …

    Notes de l’épisode

    This episode examines how to loop a transformer's layers, meaning reuse the same weights several times per token, depending on whether model size, training compute, or decoding speed is the scarce resource. It compares Full-Stack, Middle-Stack, and Per-Layer looping. Full-Stack gives the best quality at fixed parameters, Middle-Stack gives the best return at fixed training FLOPs, and gains flatten or turn unstable past about 16 loops. The 32-loop evidence comes only from the 70M and 100M models. Small per-loop low-rank adapters add only 1 to 4 percent more parameters and help more as loop counts grow, but at comparable FLOPs, doubling true depth still reaches lower loss. The discussion then turns to the decoding-latency problem, where each loop adds a sequential pass. It covers the Parallel Loop Transformer, which uses the previous token's state to run loops in parallel, and why that shift compounds prediction errors and causes magnitude mismatches. It also covers a proposed fix that re-injects the current token's embedding through normalization and a gate. Listeners get a practical guide to matching a looping design to a real resource constraint, with caveats about how far the evidence scales. Sources: 1. How to Loop Your Transformer: Loop Scaling, Specialization, and Parallel Inference — Chanakya Ekbote, Artidoro Pagnoni, Devendra Singh Sachan, Shruti Bhosale, Mike Lewis, Maryam Fazel-Zarandi, Shashank Rajput, Sujan Kumar Gonugondla, 2026 http://arxiv.org/abs/2610.40325v1 2. Universal Transformers — Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, Łukasz Kaiser, 2018 https://scholar.google.com/scholar?q=Universal+Transformers 3. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach — Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, et al., 2025 https://scholar.google.com/scholar?q=Scaling+up+Test-Time+Compute+with+Latent+Reasoning%3A+A+Recurrent+Depth+Approach 4. Reasoning with Latent Thoughts: On the Power of Looped Transformers — Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, Sashank Reddi, 2025 https://scholar.google.com/scholar?q=Reasoning+with+Latent+Thoughts%3A+On+the+Power+of+Looped+Transformers 5. Parallel Loop Transformer (PLT) — Wu et al. (full title and author list not verified; cited in the paper as Wu et al., 2025), 2025 https://scholar.google.com/scholar?q=Parallel+Loop+Transformer+%28PLT%29 6. LoRA: Low-Rank Adaptation of Large Language Models — Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, 2021 https://scholar.google.com/scholar?q=LoRA%3A+Low-Rank+Adaptation+of+Large+Language+Models 7. Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA — Sangmin Bae, Adam Fisch, Hrayr Harutyunyan, Ziwei Ji, Seungyeon Kim, Tal Schuster, 2024/2025 (cited as Bae et al., 2025 in the paper; verify the exact venue year) https://scholar.google.com/scholar?q=Relaxed+Recursive+Transformers%3A+Effective+Parameter+Sharing+with+Layer-wise+LoRA 8. Fast Inference from Transformers via Speculative Decoding — Yaniv Leviathan, Matan Kalman, Yossi Matias, 2023 https://scholar.google.com/scholar?q=Fast+Inference+from+Transformers+via+Speculative+Decoding 9. Efficient Memory Management for Large Language Model Serving with PagedAttention — Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, Ion Stoica, 2023 https://scholar.google.com/scholar?q=Efficient+Memory+Management+for+Large+Language+Model+Serving+with+PagedAttention 10. Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads — Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Tri Dao, 2024 https://scholar.google.com/scholar?q=Medusa%3A+Simple+LLM+Inference+Acceleration+Framework+with+Multiple+Decoding+Heads 11. Parallel Loop Transformer for Efficient Test-Time Computation Scaling — Bohong Wu, Mengzhao Chen, Xiang Luo, et al., 2025 https://scholar.google.com/scholar?q=Parallel+Loop+Transformer+for+Efficient+Test-Time+Computation+Scaling 12. Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation — Sangmin Bae, Yujin Kim, Reza Bayat, et al., 2025/2026 https://scholar.google.com/scholar?q=Mixture-of-Recursions%3A+Learning+Dynamic+Recursive+Depths+for+Adaptive+Token-Level+Computation 13. Parcae: Scaling Laws for Stable Looped Language Models — Hayden Prairie, Zachary Novack, Taylor Berg-Kirkpatrick, Daniel Y. Fu, 2026 https://scholar.google.com/scholar?q=Parcae%3A+Scaling+Laws+for+Stable+Looped+Language+Models 14. How Much Is One Recurrence Worth? Iso-Depth Scaling Laws for Looped Language Models — Kristian Schwethelm, Daniel Rueckert, Georgios Kaissis, 2026 https://scholar.google.com/scholar?q=How+Much+Is+One+Recurrence+Worth%3F+Iso-Depth+Scaling+Laws+for+Looped+Language+Models 15. LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling — Jian Yang, Shawn Guo, Wei Zhang, et al., 2026 https://scholar.google.com/scholar?q=LoopCoder-v2%3A+Only+Loop+Once+for+Efficient+Test-Time+Computation+Scaling 16. Hyperloop Transformers — Abbas Zeitoun, Lucas Torroba-Hennigen, Yoon Kim, 2026 https://scholar.google.com/scholar?q=Hyperloop+Transformers 17. T2MLR: Transformer with Temporal Middle-Layer Recurrence — Ziyang Cai, Xingyu Zhu, Yihe Dong, Yinghui He, Sanjeev Arora, 2026 https://scholar.google.com/scholar?q=T2MLR%3A+Transformer+with+Temporal+Middle-Layer+Recurrence 18. Scaling Latent Reasoning via Looped Language Models (Ouro) — Rui-Jie Zhu, Zixuan Wang, Kai Hua, et al., 2026 https://scholar.google.com/scholar?q=Scaling+Latent+Reasoning+via+Looped+Language+Models+%28Ouro%29 19. Staircase Attention for Recurrent Processing of Sequences — Da Ju, Stephen Roller, Sainbayar Sukhbaatar, Jason E. Weston, 2022 https://scholar.google.com/scholar?q=Staircase+Attention+for+Recurrent+Processing+of+Sequences Interactive Visualization: How to Loop Your Transformer: Scaling, Specialization, Parallel Inference

    Transcription

15. Test-Time Training via Context Distillation: Learning What to Remember

    2 novembre 2026À suivre

    This episode examines "Learning What to Remember: Test-Time Training via Context Distillation" (TTCD), which asks what a model should write into a fixed-size memory when it can't keep everything. It starts with …

    Notes de l’épisode

    This episode examines "Learning What to Remember: Test-Time Training via Context Distillation" (TTCD), which asks what a model should write into a fixed-size memory when it can't keep everything. It starts with why linear-attention and recurrent models like DeltaNet and Gated DeltaNet need lossy bounded state. It then traces the progression of test-time training from the delta rule through TTT layers, Titans, LaCT, and In-Place TTT, which reuses a transformer's MLP down-projection as fast weights. The paper's method runs the same frozen backbone twice, once with a long sliding window as teacher and once with a short window as student. It writes the residual between their MLP outputs into the fast weights through an outer-product update, so the information outlives the short window. The discussion stays skeptical. It asks whether the residual really captures remote history or just a modest window extension, whether the distillation loss actually shapes the writes, and why the closest next-token competitor, TTT-E2E, is left out of the comparisons. Sources: 1. Learning What to Remember: Test-Time Training via Context Distillation — Zixuan Wang, Xingyu Dang, Rui-Jie Zhu, Zixin Wen, Hengyu Fu, Wenhao Chai, Jason D. Lee, 2026 http://arxiv.org/abs/2608.01672 2. Learning to (Learn at Test Time): RNNs with Expressive Hidden States — Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, Carlos Guestrin, 2024 (ICML 2025 version also circulated) https://scholar.google.com/scholar?q=Learning+to+%28Learn+at+Test+Time%29%3A+RNNs+with+Expressive+Hidden+States 3. Gated Delta Networks: Improving Mamba2 with Delta Rule — Songlin Yang, Jan Kautz, Ali Hatamizadeh, 2024 (ICLR 2025) https://scholar.google.com/scholar?q=Gated+Delta+Networks%3A+Improving+Mamba2+with+Delta+Rule 4. Titans: Learning to Memorize at Test Time — Ali Behrouz, Peilin Zhong, Vahab Mirrokni, 2024 https://scholar.google.com/scholar?q=Titans%3A+Learning+to+Memorize+at+Test+Time 5. A General Method for Test-Time Training / Learning by Distilling Context — Snell, Klein, Zhong (Learning by Distilling Context); Askell et al. (A General Language Assistant as a Laboratory for Alignment), 2022 (Snell et al.); 2021 (Askell et al.) https://scholar.google.com/scholar?q=A+General+Method+for+Test-Time+Training+%2F+Learning+by+Distilling+Context 6. The Surprising Effectiveness of Test-Time Training for Few-Shot Learning — Ekin Akyürek, Mehul Damani, Linlu Qiu, Han Guo, Yoon Kim, Jacob Andreas, 2024 https://scholar.google.com/scholar?q=The+Surprising+Effectiveness+of+Test-Time+Training+for+Few-Shot+Learning 7. End-to-End Test-Time Training for Long Context (TTT-E2E) — Arnuv Tandon, Karan Dalal, Xinhao Li, et al., 2025 https://scholar.google.com/scholar?q=End-to-End+Test-Time+Training+for+Long+Context+%28TTT-E2E%29 8. In-Place Test-Time Training — Guhao Feng, Shengjie Luo, Kai Hua, Ge Zhang, Wenhao Huang, Di He, Tianle Cai, 2026 https://scholar.google.com/scholar?q=In-Place+Test-Time+Training 9. Test-Time Training Done Right (LaCT) — Tianyuan Zhang, Sai Bi, Yicong Hong, et al., 2025 https://scholar.google.com/scholar?q=Test-Time+Training+Done+Right+%28LaCT%29 10. Learning by Distilling Context — Charlie Snell, Dan Klein, Ruiqi Zhong, 2022 https://scholar.google.com/scholar?q=Learning+by+Distilling+Context 11. Cartridges: Lightweight and general-purpose long context representations via self-study — Sabri Eyuboglu, Ryan Ehrlich, Simran Arora, et al., 2025 https://scholar.google.com/scholar?q=Cartridges%3A+Lightweight+and+general-purpose+long+context+representations+via+self-study 12. Kimi Linear: An Expressive, Efficient Attention Architecture — Kimi Team, Yu Zhang, et al., 2025 https://scholar.google.com/scholar?q=Kimi+Linear%3A+An+Expressive%2C+Efficient+Attention+Architecture 13. Linear Transformers Are Secretly Fast Weight Programmers — Imanol Schlag, Kazuki Irie, Jürgen Schmidhuber, 2021 https://scholar.google.com/scholar?q=Linear+Transformers+Are+Secretly+Fast+Weight+Programmers 14. Infini-attention: Leave No Context Behind — Tsendsuren Munkhdalai, Manaal Faruqui, Siddharth Gopal, 2024 https://scholar.google.com/scholar?q=Infini-attention%3A+Leave+No+Context+Behind 15. RULER: What's the Real Context Size of Your Long-Context Language Models? — Cheng-Ping Hsieh, et al., 2024 https://scholar.google.com/scholar?q=RULER%3A+What%27s+the+Real+Context+Size+of+Your+Long-Context+Language+Models%3F 16. Mechanistic Design and Scaling of Hybrid Architectures (MAD) — Michael Poli, et al., 2024 https://scholar.google.com/scholar?q=Mechanistic+Design+and+Scaling+of+Hybrid+Architectures+%28MAD%29 Interactive Visualization: Test-Time Training via Context Distillation: Learning What to Remember

    Transcription

16. Speculative Decoding Surprisingly Excels on Sparse MoE Models

    1 novembre 2026À suivre

    This episode examines why speculative decoding, long thought to give little benefit for sparse mixture-of-experts models at batch size one, actually helps most in that regime for many newer models. It covers …

    Notes de l’épisode

    This episode examines why speculative decoding, long thought to give little benefit for sparse mixture-of-experts models at batch size one, actually helps most in that regime for many newer models. It covers ten MoE models from 30 billion to a trillion parameters, and explains how a drafter proposes tokens that the target model verifies in a single pass. It then contrasts coarse-grained experts (Mixtral-style) with fine-grained ones (Qwen3, Kimi-K2.5), where each token touches only a tiny slice of the layer. The discussion challenges the bandwidth-bound cost model from prior work, which predicts a speedup curve that rises, peaks at mid-sized batches, then falls. The authors point to two missing factors: adjacent tokens routing to overlapping experts, and real GPU execution dynamics. The episode also explains target efficiency, a metric that isolates verification cost from draft acceptance, and outlines the vLLM benchmarking setup on H100 and B200 GPUs. Listeners who care about low-latency serving for agents, test-time scaling, or RL rollouts will find it useful for deciding when speculative decoding is worth enabling. Sources: 1. Why Speculative Decoding Works Better Than Predicted on Sparse MoE Models — Ekagra Ranjan, Komal K. Teru, Bharat Venkitesh, Acyr Locatelli, 2026 http://arxiv.org/abs/2610.38981 2. Fast Inference from Transformers via Speculative Decoding — Yaniv Leviathan, Matan Kalman, Yossi Matias, 2022 (arXiv); ICML 2023 https://scholar.google.com/scholar?q=Fast+Inference+from+Transformers+via+Speculative+Decoding 3. Accelerating Large Language Model Decoding with Speculative Sampling — Charlie Chen, Sebastian Borgeaud, Geoffrey Irving, Jean-Baptiste Lespiau, Laurent Sifre, John Jumper, 2023 https://scholar.google.com/scholar?q=Accelerating+Large+Language+Model+Decoding+with+Speculative+Sampling 4. MoESD: Unveil Speculative Decoding's Potential for Accelerating Sparse MoE — Zongle Huang, Lei Zhu, Zongyuan Zhan, Ting Hu, Weikai Mao, Xianzhi Yu, Yongpan Liu, Tianyu Zhang (verify author list), 2025 https://scholar.google.com/scholar?q=MoESD%3A+Unveil+Speculative+Decoding%27s+Potential+for+Accelerating+Sparse+MoE 5. EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test — Yuhui Li, Fangyun Wei, Chao Zhang, Hongyang Zhang, 2025 https://scholar.google.com/scholar?q=EAGLE-3%3A+Scaling+up+Inference+Acceleration+of+Large+Language+Models+via+Training-Time+Test 6. Exploiting Inter-Layer Expert Affinity for Accelerating Mixture-of-Experts Model Inference — Jinghan Yao, Quentin Anthony, Aamir Shafi, Hari Subramoni, Dhabaleswar K. Panda, 2024 https://scholar.google.com/scholar?q=Exploiting+Inter-Layer+Expert+Affinity+for+Accelerating+Mixture-of-Experts+Model+Inference 7. Fast Inference of Mixture-of-Experts Language Models with Offloading — Artyom Eliseev, Denis Mazur, 2023 https://scholar.google.com/scholar?q=Fast+Inference+of+Mixture-of-Experts+Language+Models+with+Offloading 8. DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models — Damai Dai, Chengqi Deng, Chenggang Zhao, and others (DeepSeek-AI), 2024 https://scholar.google.com/scholar?q=DeepSeekMoE%3A+Towards+Ultimate+Expert+Specialization+in+Mixture-of-Experts+Language+Models 9. Mixtral of Experts — Albert Q. Jiang and others (Mistral AI), 2024 https://scholar.google.com/scholar?q=Mixtral+of+Experts 10. On a Least Squares Adjustment of a Sampled Frequency Table When the Expected Marginal Totals are Known — W. Edwards Deming, Frederick F. Stephan, 1940 https://scholar.google.com/scholar?q=On+a+Least+Squares+Adjustment+of+a+Sampled+Frequency+Table+When+the+Expected+Marginal+Totals+are+Known 11. I-Divergence Geometry of Probability Distributions and Minimization Problems — Imre Csiszár, 1975 https://scholar.google.com/scholar?q=I-Divergence+Geometry+of+Probability+Distributions+and+Minimization+Problems 12. Information Theory and Statistical Mechanics — E. T. Jaynes, 1957 https://scholar.google.com/scholar?q=Information+Theory+and+Statistical+Mechanics 13. Sinkhorn Distances: Lightspeed Computation of Optimal Transport — Marco Cuturi, 2013 https://scholar.google.com/scholar?q=Sinkhorn+Distances%3A+Lightspeed+Computation+of+Optimal+Transport 14. Roofline: An Insightful Visual Performance Model for Multicore Architectures — Samuel Williams, Andrew Waterman, David Patterson, 2009 https://scholar.google.com/scholar?q=Roofline%3A+An+Insightful+Visual+Performance+Model+for+Multicore+Architectures 15. Efficiently Scaling Transformer Inference — Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Anselm Levskaya, Jonathan Heek, Kefan Xiao, Shivani Agrawal, Jeff Dean, 2022 https://scholar.google.com/scholar?q=Efficiently+Scaling+Transformer+Inference 16. MegaBlocks: Efficient Sparse Training with Mixture-of-Experts — Trevor Gale, Deepak Narayanan, Cliff Young, Matei Zaharia, 2023 https://scholar.google.com/scholar?q=MegaBlocks%3A+Efficient+Sparse+Training+with+Mixture-of-Experts 17. Dissecting the NVIDIA Hopper Architecture through Microbenchmarking and Multiple Level Analysis — Weile Luo, Ruibo Fan, Zeyu Li, Dayou Du, Qiang Wang, Xiaowen Chu, 2025 https://scholar.google.com/scholar?q=Dissecting+the+NVIDIA+Hopper+Architecture+through+Microbenchmarking+and+Multiple+Level+Analysis 18. Utility-driven Speculative Decoding for Mixture-of-Experts — Saxena, Tsai, Taneja, Jaleel, Qureshi, 2025 https://scholar.google.com/scholar?q=Utility-driven+Speculative+Decoding+for+Mixture-of-Experts 19. MagicDec: Breaking the Latency-Throughput Tradeoff for Long Context Generation with Speculative Decoding — Sadhukhan et al., 2025 https://scholar.google.com/scholar?q=MagicDec%3A+Breaking+the+Latency-Throughput+Tradeoff+for+Long+Context+Generation+with+Speculative+Decoding 20. ExpertFlow: Efficient MoE Inference via Predictive Expert Caching and Token Scheduling — He et al., 2026 https://scholar.google.com/scholar?q=ExpertFlow%3A+Efficient+MoE+Inference+via+Predictive+Expert+Caching+and+Token+Scheduling 21. Probing Semantic Routing in Large Mixture-of-Expert Models — Olson et al., 2025 https://scholar.google.com/scholar?q=Probing+Semantic+Routing+in+Large+Mixture-of-Expert+Models 22. On the Spatial Structure of Mixture-of-Experts in Transformers — Bershatsky, Oseledets, 2025 https://scholar.google.com/scholar?q=On+the+Spatial+Structure+of+Mixture-of-Experts+in+Transformers 23. Scaling Laws for Fine-Grained Mixture of Experts — Ludziejewski et al., 2024 https://scholar.google.com/scholar?q=Scaling+Laws+for+Fine-Grained+Mixture+of+Experts 24. On a Least Squares Adjustment of a Sampled Frequency Table When the Expected Marginal Totals are Known / I-divergence Geometry of Probability Distributions and Minimization Problems — Deming & Stephan (1940); Csiszár (1975), 1940 / 1975 https://scholar.google.com/scholar?q=On+a+Least+Squares+Adjustment+of+a+Sampled+Frequency+Table+When+the+Expected+Marginal+Totals+are+Known+%2F+I-divergence+Geometry+of+Probability+Distributions+and+Minimization+Problems Interactive Visualization: Speculative Decoding Surprisingly Excels on Sparse MoE Models

    Transcription

17. FlatCake: Flat KVCache Storage Beats Tiered Memory Hierarchies

    1 novembre 2026À suivre

    This episode examines FlatCake, a proposed KV cache storage architecture for LLM serving that drops the usual GPU, host DRAM, and NVMe SSD hierarchy. Instead of migrating blocks between tiers, it decides at …

    Notes de l’épisode

    This episode examines FlatCake, a proposed KV cache storage architecture for LLM serving that drops the usual GPU, host DRAM, and NVMe SSD hierarchy. Instead of migrating blocks between tiers, it decides at write time whether each block lives in DRAM or on SSD, and it never moves the block afterward. The discussion explains why the authors think tiering breaks down. Reuse is bimodal: in several conversational and API traces, most blocks are used only once. In a Claude Code agentic trace, sessions are so large that a few concurrent ones overflow DRAM. Staging buffers also run out under load, and blocks written in eviction order end up scattered across the drive. The hosts walk through the proposed fix, which lays each session out contiguously on the SSD so a restore becomes a few large reads, and they note that GPUDirect Storage fixes the bounce buffer but not the scatter. They also point out the cost of the approach: placement relies on hand-tuned thresholds with no published values, and a wrong placement choice sticks. Listeners interested in serving infrastructure and storage layout get a critical look at how much the paper's experiments actually test its claims. Sources: 1. FlatCake: A Flat KVCache Storage Architecture for LLM Serving — Xin Huang, Teng Ma, Zhuoran Li, Yibo Zhao, Zhaoshi Li, Qianchao Zhu, Wei Yang, Chenxi Wang, Tao Xie, Youwei Zhuo, 2026 http://arxiv.org/abs/2610.38355 2. Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving — Ruoyu Qin, Zheming Li, Weiran He, Mingxing Zhang, Yongwei Wu, Weimin Zheng, Xinran Xu (Moonshot AI / Tsinghua), 2024 (FAST 2025) https://scholar.google.com/scholar?q=Mooncake%3A+A+KVCache-centric+Disaggregated+Architecture+for+LLM+Serving 3. Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention (AttentionStore) — Bin Gao, Zhuomin He, Puru Sharma, Qingxuan Kang, Djordje Jevdjic, Junbo Deng, Xingkun Yang, Zhou Yu, Pengfei Zuo, 2024 (USENIX ATC 2024) https://scholar.google.com/scholar?q=Cost-Efficient+Large+Language+Model+Serving+for+Multi-turn+Conversations+with+CachedAttention+%28AttentionStore%29 4. LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference — Yihua Cheng, Yuhan Liu, Jiayi Yao, Yuwei An, Xiaokun Chen, Shaoting Feng, Yuyang Huang, Samuel Shen, Kuntai Du, Junchen Jiang et al., 2025 https://scholar.google.com/scholar?q=LMCache%3A+An+Efficient+KV+Cache+Layer+for+Enterprise-Scale+LLM+Inference 5. FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU — Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, Beidi Chen, Percy Liang, Christopher Ré, Ion Stoica, Ce Zhang, 2023 (ICML 2023) https://scholar.google.com/scholar?q=FlexGen%3A+High-Throughput+Generative+Inference+of+Large+Language+Models+with+a+Single+GPU 6. ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning — Samyam Rajbhandari, Olatunji Ruwase, Jeff Rasley, Shaden Smith, Yuxiong He, 2021 (SC 2021) https://scholar.google.com/scholar?q=ZeRO-Infinity%3A+Breaking+the+GPU+Memory+Wall+for+Extreme+Scale+Deep+Learning 7. GPU-Initiated On-Demand High-Throughput Storage Access in the BaM System Architecture — Zaid Qureshi, Vikram Sharma Mailthody, Isaac Gelado, Seung Won Min, Amna Masood, Jeongmin Park, Jinjun Xiong, C. J. Newburn, Dmitri Vainbrand, I-Hsin Chung, Michael Garland, William Dally, Wen-mei Hwu, 2023 (ASPLOS 2023) https://scholar.google.com/scholar?q=GPU-Initiated+On-Demand+High-Throughput+Storage+Access+in+the+BaM+System+Architecture 8. GIDS: Accelerating Sampling and Aggregation Operations in GNN Frameworks with GPU Initiated Direct Storage Accesses — Jeongmin Brian Park, Vikram Sharma Mailthody, Zaid Qureshi, Wen-mei Hwu, 2024 (PVLDB 2024) https://scholar.google.com/scholar?q=GIDS%3A+Accelerating+Sampling+and+Aggregation+Operations+in+GNN+Frameworks+with+GPU+Initiated+Direct+Storage+Accesses 9. KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider — Jiahao Wang, Jinbo Han, Xingda Wei, Sijie Shen, Dingyan Zhang, Chenguang Fang, Rong Chen, Wenyuan Yu, Haibo Chen, 2025 (USENIX ATC 2025) https://scholar.google.com/scholar?q=KVCache+Cache+in+the+Wild%3A+Characterizing+and+Optimizing+KVCache+Cache+at+a+Large+Cloud+Provider 10. Preble: Efficient Distributed Prompt Scheduling for LLM Serving — Vikranth Srivastava, Zijian He, Shuowei Jin (Yingxuan Jin?), Yilong Zhao, Yibo Zhu, Ion Stoica, Hao Zhang and others (UCSD / Berkeley), 2024 (ICLR 2025) https://scholar.google.com/scholar?q=Preble%3A+Efficient+Distributed+Prompt+Scheduling+for+LLM+Serving 11. FIFO Queues are All You Need for Cache Eviction — Juncheng Yang, Yazhuo Zhang, Ziyue Qiu, Yao Yue, Rashmi Vinayak, 2023 (SOSP 2023) https://scholar.google.com/scholar?q=FIFO+Queues+are+All+You+Need+for+Cache+Eviction 12. CacheGen: KV Cache Compression and Streaming for Fast Large Language Model Serving — Yuhan Liu, Hanchen Li, Yihua Cheng, et al., 2024 https://scholar.google.com/scholar?q=CacheGen%3A+KV+Cache+Compression+and+Streaming+for+Fast+Large+Language+Model+Serving 13. CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion — Jiayi Yao, Hanchen Li, Yuhan Liu, et al., 2025 https://scholar.google.com/scholar?q=CacheBlend%3A+Fast+Large+Language+Model+Serving+for+RAG+with+Cached+Knowledge+Fusion 14. AttentionStore / CachedAttention: Cost-Efficient LLM Serving for Multi-turn Conversations — Bin Gao, Zhuomin He, Puru Sharma, et al., 2024 https://scholar.google.com/scholar?q=AttentionStore+%2F+CachedAttention%3A+Cost-Efficient+LLM+Serving+for+Multi-turn+Conversations 15. Managing Scalable Direct Storage Accesses for GPUs with GoFS — Shaobo Li, Yirui Eric Zhou, Yuqi Xue, Yuan Xu, Jian Huang, 2025 https://scholar.google.com/scholar?q=Managing+Scalable+Direct+Storage+Accesses+for+GPUs+with+GoFS 16. Bidaw: Enhancing Key-Value Caching for Interactive LLM Serving via Bidirectional Computation–Storage Awareness — Shipeng Hu, Guangyan Zhang, et al., 2026 https://scholar.google.com/scholar?q=Bidaw%3A+Enhancing+Key-Value+Caching+for+Interactive+LLM+Serving+via+Bidirectional+Computation%E2%80%93Storage+Awareness 17. Tutti: GPU-initiated io_uring queues for KV storage — Cited as [28] in the paper, 2025–2026 https://scholar.google.com/scholar?q=Tutti%3A+GPU-initiated+io_uring+queues+for+KV+storage Interactive Visualization: FlatCake: Flat KVCache Storage Beats Tiered Memory Hierarchies

    Transcription

18. Agent Memory Costs: Characterizing Write-Time Overhead in Long-Horizon Workloads

    1 novembre 2026À suivre

    This episode examines a paper that profiles ten agent-memory systems as systems workloads and splits their cost into construction (the write path), retrieval and generation, rather than comparing them on …

    Notes de l’épisode

    This episode examines a paper that profiles ten agent-memory systems as systems workloads and splits their cost into construction (the write path), retrieval and generation, rather than comparing them on accuracy alone. It covers why pasting full history into a long context fails: prefill cost grows quadratically, prefix caching breaks across sessions, and recall degrades for material in the middle of long inputs. It then explains how agent memory differs from RAG and walks through a seven-stage pipeline and four construction styles, from none to deterministic chunking, LLM-mediated fact extraction and agentic control. The ten systems, including BM25, GraphRAG, Mem0, Letta and MIRIX, are placed along a gradient of increasing LLM involvement at write time. The discussion also lays out the measurement setup (phase-tagged calls aligned with GPU telemetry on a single H100 with Qwen3 models, plus remote OpenAI models) and previews the argument that construction is the hidden bill. Listeners get a concrete view of where long-lived agents spend compute and energy, and why that matters for scheduling. Sources: 1. Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads — Yasmine Omri, Ziyu Gan, Zachary Broveak, Robin Geens, Zexue He, Alex Pentland, Marian Verhelst, Tsachy Weissman, Thierry Tambe, 2026 http://arxiv.org/abs/2606.06448 2. MemGPT: Towards LLMs as Operating Systems — Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez, 2023 https://scholar.google.com/scholar?q=MemGPT%3A+Towards+LLMs+as+Operating+Systems 3. Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory — Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, Deshraj Yadav, 2025 https://scholar.google.com/scholar?q=Mem0%3A+Building+Production-Ready+AI+Agents+with+Scalable+Long-Term+Memory 4. Evaluating Very Long-Term Conversational Memory of LLM Agents (LoCoMo) — Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, Yuwei Fang, 2024 https://scholar.google.com/scholar?q=Evaluating+Very+Long-Term+Conversational+Memory+of+LLM+Agents+%28LoCoMo%29 5. Sleep-time Compute: Beyond Inference Scaling at Test-time — Kevin Lin, Charlie Snell, Yu Wang, Charles Packer, Sarah Wooders, Ion Stoica, Joseph E. Gonzalez, 2025 https://scholar.google.com/scholar?q=Sleep-time+Compute%3A+Beyond+Inference+Scaling+at+Test-time 6. Zep: A Temporal Knowledge Graph Architecture for Agent Memory — Preston Rasmussen, Pavel Paliychuk, Travis Beauvais, Jack Ryan, Daniel Chalef, 2025 https://scholar.google.com/scholar?q=Zep%3A+A+Temporal+Knowledge+Graph+Architecture+for+Agent+Memory 7. LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference — Yihua Cheng et al., 2025 https://scholar.google.com/scholar?q=LMCache%3A+An+Efficient+KV+Cache+Layer+for+Enterprise-Scale+LLM+Inference 8. Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving — Ruoyu Qin et al., 2024 https://scholar.google.com/scholar?q=Mooncake%3A+A+KVCache-centric+Disaggregated+Architecture+for+LLM+Serving 9. Sarathi-Serve: Taming Throughput-Latency Tradeoff in LLM Inference with Chunked Prefills — Amey Agrawal et al., 2024 https://scholar.google.com/scholar?q=Sarathi-Serve%3A+Taming+Throughput-Latency+Tradeoff+in+LLM+Inference+with+Chunked+Prefills 10. Zep, LightMem, and other efficiency-oriented memory systems (e.g., LightMem: Lightweight and Efficient Memory-Augmented Generation) — Various, 2025 https://scholar.google.com/scholar?q=Zep%2C+LightMem%2C+and+other+efficiency-oriented+memory+systems+%28e.g.%2C+LightMem%3A+Lightweight+and+Efficient+Memory-Augmented+Generation%29 11. Cartridges: Lightweight and General-Purpose Long Context Representations via Self-Study — Sabri Eyuboglu et al., 2025 https://scholar.google.com/scholar?q=Cartridges%3A+Lightweight+and+General-Purpose+Long+Context+Representations+via+Self-Study 12. Efficient Memory Management for Large Language Model Serving with PagedAttention (vLLM) — Woosuk Kwon et al., 2023 https://scholar.google.com/scholar?q=Efficient+Memory+Management+for+Large+Language+Model+Serving+with+PagedAttention+%28vLLM%29 13. Large-scale cluster management at Google with Borg — Abhishek Verma et al., 2015 https://scholar.google.com/scholar?q=Large-scale+cluster+management+at+Google+with+Borg Interactive Visualization: Agent Memory Costs: Characterizing Write-Time Overhead in Long-Horizon Workloads

    Transcription

19. Investor's Last Exam: Benchmarking AI Equity Research Without Future Leakage

    1 novembre 2026À suivre

    This episode examines a paper that builds a benchmark for testing whether AI agents can do equity investment research without seeing the future. It centers on a Tesla task whose search used a "before March 24 …

    Notes de l’épisode

    This episode examines a paper that builds a benchmark for testing whether AI agents can do equity investment research without seeing the future. It centers on a Tesla task whose search used a "before March 24, 2026" date filter. The filter still returned a July page reporting realized Q2 deliveries of 480,126, because pages can keep an old publication date after their content is updated. The benchmark's fix is version-level availability: every document revision is stored separately and is eligible only if it predates the decision time. The database holds about 3.4 million versions, 1.52 TB, and every tool call is checked against it. The discussion explains why finance lacks a pass/fail oracle like SWE-bench. It describes the expert rubrics and dependency graphs that grade the chain from evidence to assumptions to judgment. It also covers the three yardsticks of answer quality, forecast outcome and simulated return, which the paper says rank models differently. The hosts separate retrieval leakage from "parametric leakage," meaning outcomes already in a model's pretraining weights, and tie that to prior work on lookahead bias in finance. The episode closes by introducing the 86 buy-side-style tasks, the 11 agent systems and the shared six-tool setup. Sources: 1. Investor's Last Exam: Can AI Agents Do Equity Investment Without Seeing the Future? — Huangfei Jiang, Anning Hu, Kai Sheng, Yuchen Cai, Rui Ge, Xiangyu Kong, Jianghao Lin, Shuzi Chen, Jiyi Weng, Junchi Yan, 2026 http://arxiv.org/abs/2610.38342 2. FinanceBench: A New Benchmark for Financial Question Answering — Pranab Islam, Anand Kannappan, Douwe Kiela, Rebecca Qian, Nino Scherrer, Bertie Vidgen, 2023 https://scholar.google.com/scholar?q=FinanceBench%3A+A+New+Benchmark+for+Financial+Question+Answering 3. BloombergGPT: A Large Language Model for Finance — Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, Gideon Mann, 2023 https://scholar.google.com/scholar?q=BloombergGPT%3A+A+Large+Language+Model+for+Finance 4. TradingAgents: Multi-Agents LLM Financial Trading Framework — Yijia Xiao, Edward Sun, Di Luo, Wei Wang, 2024 https://scholar.google.com/scholar?q=TradingAgents%3A+Multi-Agents+LLM+Financial+Trading+Framework 5. Financial Statement Analysis with Large Language Models — Alex G. Kim, Maximilian Muhn, Valeri V. Nikolaev, 2024 https://scholar.google.com/scholar?q=Financial+Statement+Analysis+with+Large+Language+Models 6. Assessing Look-Ahead Bias in Stock Return Predictions Generated By GPT Sentiment Analysis — Paul Glasserman, Caden Lin, 2023 https://scholar.google.com/scholar?q=Assessing+Look-Ahead+Bias+in+Stock+Return+Predictions+Generated+By+GPT+Sentiment+Analysis 7. Lookahead Bias in Pretrained Language Models — Suproteem Sarkar, Keyon Vafa, 2024 https://scholar.google.com/scholar?q=Lookahead+Bias+in+Pretrained+Language+Models 8. Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance — David H. Bailey, Jonathan M. Borwein, Marcos López de Prado, Qiji Jim Zhu, 2014 https://scholar.google.com/scholar?q=Pseudo-Mathematics+and+Financial+Charlatanism%3A+The+Effects+of+Backtest+Overfitting+on+Out-of-Sample+Performance 9. Advances in Financial Machine Learning — Marcos López de Prado, 2018 https://scholar.google.com/scholar?q=Advances+in+Financial+Machine+Learning 10. HealthBench: Evaluating Large Language Models Towards Improved Human Health — Rahul K. Arora et al. (OpenAI), 2025 https://scholar.google.com/scholar?q=HealthBench%3A+Evaluating+Large+Language+Models+Towards+Improved+Human+Health 11. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena — Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, et al., 2023 https://scholar.google.com/scholar?q=Judging+LLM-as-a-Judge+with+MT-Bench+and+Chatbot+Arena 12. Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains — Anisha Gunjal et al., 2025 https://scholar.google.com/scholar?q=Rubrics+as+Rewards%3A+Reinforcement+Learning+Beyond+Verifiable+Domains 13. Humanity's Last Exam — Long Phan et al. (Center for AI Safety and Scale AI), 2025 https://scholar.google.com/scholar?q=Humanity%27s+Last+Exam 14. Approaching Human-Level Forecasting with Language Models — Danny Halawi, Fred Zhang, Chen Yueh-Han, Jacob Steinhardt, 2024 https://scholar.google.com/scholar?q=Approaching+Human-Level+Forecasting+with+Language+Models 15. ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities — Ezra Karger, Houtan Bastani, Chen Yueh-Han, Zachary Jacobs, Danny Halawi, Fred Zhang, Philip E. Tetlock, 2024 https://scholar.google.com/scholar?q=ForecastBench%3A+A+Dynamic+Benchmark+of+AI+Forecasting+Capabilities 16. Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models — Alejandro Lopez-Lira, Yuehua Tang, 2023 https://scholar.google.com/scholar?q=Can+ChatGPT+Forecast+Stock+Price+Movements%3F+Return+Predictability+and+Large+Language+Models 17. Chronologically Consistent Large Language Models — Songrun He, Linying Lv, Asaf Manela, Jimmy Wu, 2025 https://scholar.google.com/scholar?q=Chronologically+Consistent+Large+Language+Models 18. Look-Ahead-Bench: A standardized benchmark of look-ahead bias in point-in-time LLMs for finance — Mostapha Benhenda, 2026 https://scholar.google.com/scholar?q=Look-Ahead-Bench%3A+A+standardized+benchmark+of+look-ahead+bias+in+point-in-time+LLMs+for+finance 19. Pitfalls in Evaluating Language Model Forecasters — Daniel Paleka, Shashwat Goel, Jonas Geiping, Florian Tramèr, 2025 https://scholar.google.com/scholar?q=Pitfalls+in+Evaluating+Language+Model+Forecasters 20. Temporal leakage in search-engine date-filtered web retrieval: A retrospective forecasting case study — Ali El Lahib, Ying-Jieh Xia, Zehan Li, Yuxuan Wang, Xinyu Pi, 2026 https://scholar.google.com/scholar?q=Temporal+leakage+in+search-engine+date-filtered+web+retrieval%3A+A+retrospective+forecasting+case+study 21. Bench to the Future: A Pastcasting Benchmark for Forecasting Agents — FutureSearch et al., 2025 https://scholar.google.com/scholar?q=Bench+to+the+Future%3A+A+Pastcasting+Benchmark+for+Forecasting+Agents 22. LLM Evaluators Recognize and Favor Their Own Generations — Arjun Panickssery, Samuel R. Bowman, Shi Feng, 2024 https://scholar.google.com/scholar?q=LLM+Evaluators+Recognize+and+Favor+Their+Own+Generations 23. GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks — Tejal Patwardhan et al., 2025 https://scholar.google.com/scholar?q=GDPval%3A+Evaluating+AI+Model+Performance+on+Real-World+Economically+Valuable+Tasks 24. Advances in Financial Machine Learning (purged and embargoed cross-validation) — Marcos López de Prado, 2018 https://scholar.google.com/scholar?q=Advances+in+Financial+Machine+Learning+%28purged+and+embargoed+cross-validation%29 25. The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality — David H. Bailey, Marcos López de Prado, 2014 https://scholar.google.com/scholar?q=The+Deflated+Sharpe+Ratio%3A+Correcting+for+Selection+Bias%2C+Backtest+Overfitting+and+Non-Normality 26. StockBench: Can LLM Agents Trade Stocks Profitably in Real-World Markets? — Yanxu Chen et al., 2025 https://scholar.google.com/scholar?q=StockBench%3A+Can+LLM+Agents+Trade+Stocks+Profitably+in+Real-World+Markets%3F Interactive Visualization: Investor's Last Exam: Benchmarking AI Equity Research Without Future Leakage

    Transcription

20. RaBitQ-SSD: Split Codes and Pipelined I/O for SSD Vector Search

    1 novembre 2026À suivre

    This episode examines RaBitQ-SSD, a system for approximate nearest neighbor search over vectors stored on SSD, which the paper says can serve ten billion vectors on one drive at over 3,000 queries per second …

    Notes de l’épisode

    This episode examines RaBitQ-SSD, a system for approximate nearest neighbor search over vectors stored on SSD, which the paper says can serve ten billion vectors on one drive at over 3,000 queries per second. It starts with the basics: why billion-scale embeddings are kept on SSD instead of DRAM, how recall at k is measured, how IVF partitions vectors into inverted lists, and why graph indexes such as DiskANN and PipeANN pay a disk read for every uncached hop. It then covers how RaBitQ's one-bit-per-dimension codes give each candidate a distance estimate with a probabilistic lower bound. That bound lets the system skip individual disk reads, and the paper reports about 38 pages read per query against roughly 870 for SPANN on Wikipedia embeddings. The hosts weigh the headline claims against a simple throughput model: once the SSD saturates at its measured 1.63 million reads per second, throughput depends on pages read per query. They also explain how the Split-RaBitQ design keeps only a prefix of each binary code in memory, rescaled so the estimate stays unbiased. Listeners get a clear account of the memory and blocking-I/O problems that make a direct port to SSD impractical. Sources: 1. RaBitQ-SSD: Split Codes and Pipelined I/O for SSD-Resident Vector Search — Yuexuan Xu, Jianyang Gao, Michael Norris, Alibek Zhakubayev, Pankaj Singh, Junjie Qi, Matthijs Douze, Cheng Long, 2026 http://arxiv.org/abs/2610.02652 2. RaBitQ: Quantizing High-Dimensional Vectors with a Theoretical Error Bound for Approximate Nearest Neighbor Search — Jianyang Gao, Cheng Long, 2024 https://scholar.google.com/scholar?q=RaBitQ%3A+Quantizing+High-Dimensional+Vectors+with+a+Theoretical+Error+Bound+for+Approximate+Nearest+Neighbor+Search 3. Practical and Asymptotically Optimal Quantization of High-Dimensional Vectors in Euclidean Space for Approximate Nearest Neighbor Search (Extended RaBitQ) — Jianyang Gao, Yutong Gou, Yuexuan Xu, Yongyi Yang, Cheng Long, Raymond Chi-Wing Wong, 2025 https://scholar.google.com/scholar?q=Practical+and+Asymptotically+Optimal+Quantization+of+High-Dimensional+Vectors+in+Euclidean+Space+for+Approximate+Nearest+Neighbor+Search+%28Extended+RaBitQ%29 4. DiskANN: Fast Accurate Billion-point Nearest Neighbor Search on a Single Node — Suhas Jayaram Subramanya, Fnu Devvrit, Harsha Vardhan Simhadri, Ravishankar Krishnaswamy, Rohan Kadekodi, 2019 https://scholar.google.com/scholar?q=DiskANN%3A+Fast+Accurate+Billion-point+Nearest+Neighbor+Search+on+a+Single+Node 5. SPANN: Highly-efficient Billion-scale Approximate Nearest Neighbor Search — Qi Chen, Bing Zhao, Haidong Wang, Mingqin Li, Chuanjie Liu, Zengzhong Li, Mao Yang, Jingdong Wang, 2021 https://scholar.google.com/scholar?q=SPANN%3A+Highly-efficient+Billion-scale+Approximate+Nearest+Neighbor+Search 6. PipeANN: Low-Latency and High-Throughput Billion-Scale Vector Search via Pipelined I/O and Compute (title as I recall it; verify) — Not confirmed from memory; verify against the citation in the paper (reference [22]), 2025 https://scholar.google.com/scholar?q=PipeANN%3A+Low-Latency+and+High-Throughput+Billion-Scale+Vector+Search+via+Pipelined+I%2FO+and+Compute+%28title+as+I+recall+it%3B+verify%29 7. PipeANN: Low-Latency Billion-Scale Vector Search via Speculative I/O Pipelining — Hao Guo, Youyou Lu, 2025 https://scholar.google.com/scholar?q=PipeANN%3A+Low-Latency+Billion-Scale+Vector+Search+via+Speculative+I%2FO+Pipelining 8. Starling: An I/O-Efficient Disk-Resident Graph Index Framework for High-Dimensional Vector Similarity Search on Data Segment — Mengzhao Wang et al., 2024 https://scholar.google.com/scholar?q=Starling%3A+An+I%2FO-Efficient+Disk-Resident+Graph+Index+Framework+for+High-Dimensional+Vector+Similarity+Search+on+Data+Segment 9. Product Quantization for Nearest Neighbor Search / Quicker ADC (FastScan) — Hervé Jégou, Matthijs Douze, Cordelia Schmid; André et al., 2011 / 2015 https://scholar.google.com/scholar?q=Product+Quantization+for+Nearest+Neighbor+Search+%2F+Quicker+ADC+%28FastScan%29 10. HELMSMAN (learned list selection with an SPDK I/O path for IVF on SSD) — As cited in the paper, 2025–2026 https://scholar.google.com/scholar?q=HELMSMAN+%28learned+list+selection+with+an+SPDK+I%2FO+path+for+IVF+on+SSD%29 11. AlayaLaser: Efficient Index Layout and Search Strategy for Large-scale High-dimensional Vector Similarity Search — As cited in the paper (ref [10]), 2025 https://scholar.google.com/scholar?q=AlayaLaser%3A+Efficient+Index+Layout+and+Search+Strategy+for+Large-scale+High-dimensional+Vector+Similarity+Search 12. VeloANN and SkipDisk — As cited in the paper (refs [73], [52]), 2025–2026 https://scholar.google.com/scholar?q=VeloANN+and+SkipDisk 13. Accelerating Large-Scale Inference with Anisotropic Vector Quantization (ScaNN) — Ruiqi Guo et al., 2020 https://scholar.google.com/scholar?q=Accelerating+Large-Scale+Inference+with+Anisotropic+Vector+Quantization+%28ScaNN%29 14. Filtered-DiskANN / FreshDiskANN — Gollapudi et al.; Singh et al., 2023 / 2021 https://scholar.google.com/scholar?q=Filtered-DiskANN+%2F+FreshDiskANN Interactive Visualization: RaBitQ-SSD: Split Codes and Pipelined I/O for SSD Vector Search

    Transcription

21. Vector Policy Optimization: Training for Diversity to Improve Test-Time Search

    31 octobre 2026À suivre

    This episode examines "Vector Policy Optimization," a proposed drop-in replacement for the GRPO advantage that trains models to produce diverse candidates rather than a single best answer, so that test-time …

    Notes de l’épisode

    This episode examines "Vector Policy Optimization," a proposed drop-in replacement for the GRPO advantage that trains models to produce diverse candidates rather than a single best answer, so that test-time search such as best-of-N and evolutionary code search has more to work with. It lays out why standard RL post-training collapses onto one high-reward mode. The evidence cited is that RL models beat base models at small k but fall behind at large k, and that KL-regularized RL is argued to collapse structurally. The method treats rewards as vectors, such as per-test-case or per-hop scores. It samples random scalarization weights from a uniform Dirichlet during training, so candidates can land at different points on the Pareto frontier. The discussion compares this with pass@k optimization, entropy bonuses and embedding-based diversity rewards, and with a multi-answer baseline that gives the model room to diversify but no incentive to do so. It also defines the evaluation metrics, pass@k, best@k and reward-space diversity. For listeners, the interest is the argument that giving up some single-shot accuracy can pay off once a search procedure sits on top of the model. Sources: 1. Vector Policy Optimization: Training for Diversity Improves Test-Time Search — Ryan Bahlous-Boldi, Isha Puri, Idan Shenfeld, Akarsh Kumar, Mehul Damani, Sebastian Risi, Omar Khattab, Zhang-Wei Hong, Pulkit Agrawal, 2026 http://arxiv.org/abs/2605.22817 2. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models — Zhihong Shao, Peiyi Wang, Qihao Zhu, et al. (DeepSeek-AI), 2024 https://scholar.google.com/scholar?q=DeepSeekMath%3A+Pushing+the+Limits+of+Mathematical+Reasoning+in+Open+Language+Models 3. Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? — Yang Yue, Zhiqi Chen, Rui Lu, et al., 2025 https://scholar.google.com/scholar?q=Does+Reinforcement+Learning+Really+Incentivize+Reasoning+Capacity+in+LLMs+Beyond+the+Base+Model%3F 4. Large Language Monkeys: Scaling Inference Compute with Repeated Sampling — Bradley Brown, Jordan Juravsky, Ryan Ehrlich, et al., 2024 https://scholar.google.com/scholar?q=Large+Language+Monkeys%3A+Scaling+Inference+Compute+with+Repeated+Sampling 5. Understanding the Effects of RLHF on LLM Generalisation and Diversity — Robert Kirk, Ishita Mediratta, Christoforos Nalmpantis, et al., 2024 https://scholar.google.com/scholar?q=Understanding+the+Effects+of+RLHF+on+LLM+Generalisation+and+Diversity 6. A Survey of Multi-Objective Sequential Decision-Making — Diederik M. Roijers, Peter Vamplew, Shimon Whiteson, Richard Dazeley, 2013 https://scholar.google.com/scholar?q=A+Survey+of+Multi-Objective+Sequential+Decision-Making 7. A Practical Guide to Multi-Objective Reinforcement Learning and Planning — Conor F. Hayes, Roxana Rădulescu, Eugenio Bargiacchi, et al., 2022 https://scholar.google.com/scholar?q=A+Practical+Guide+to+Multi-Objective+Reinforcement+Learning+and+Planning 8. A Generalized Algorithm for Multi-Objective Reinforcement Learning and Policy Adaptation — Runzhe Yang, Xingyuan Sun, Karthik Narasimhan, 2019 https://scholar.google.com/scholar?q=A+Generalized+Algorithm+for+Multi-Objective+Reinforcement+Learning+and+Policy+Adaptation 9. Rewarded Soups: Towards Pareto-Optimal Alignment by Interpolating Weights Fine-Tuned on Diverse Rewards — Alexandre Ramé, Guillaume Couairon, Corentin Dancette, et al., 2023 https://scholar.google.com/scholar?q=Rewarded+Soups%3A+Towards+Pareto-Optimal+Alignment+by+Interpolating+Weights+Fine-Tuned+on+Diverse+Rewards 10. Pass@k Policy Optimization: Solving Harder Reinforcement Learning Problems — Christian Walder, Deep Karkhanis, 2025 https://scholar.google.com/scholar?q=Pass%40k+Policy+Optimization%3A+Solving+Harder+Reinforcement+Learning+Problems 11. Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models — Yinlam Chow, Guy Tennenholtz, Izzeddin Gur, et al. (Google DeepMind), 2024 https://scholar.google.com/scholar?q=Inference-Aware+Fine-Tuning+for+Best-of-N+Sampling+in+Large+Language+Models 12. AlphaEvolve: A coding agent for scientific and algorithmic discovery — Alexander Novikov, Ngân Vũ, Marvin Eisenberger, et al. (Google DeepMind), 2025 https://scholar.google.com/scholar?q=AlphaEvolve%3A+A+coding+agent+for+scientific+and+algorithmic+discovery 13. Mathematical discoveries from program search with large language models (FunSearch) — Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, et al., 2024 https://scholar.google.com/scholar?q=Mathematical+discoveries+from+program+search+with+large+language+models+%28FunSearch%29 14. ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution — Robert Tjarko Lange, Yuki Imajuku, Edoardo Cetin (Sakana AI), 2025 https://scholar.google.com/scholar?q=ShinkaEvolve%3A+Towards+Open-Ended+and+Sample-Efficient+Program+Evolution 15. Illuminating search spaces by mapping elites (MAP-Elites) / Quality Diversity: A New Frontier for Evolutionary Computation — Jean-Baptiste Mouret and Jeff Clune (2015); Justin K. Pugh, Lisa B. Soros, Kenneth O. Stanley (2016), 2015-2016 https://scholar.google.com/scholar?q=Illuminating+search+spaces+by+mapping+elites+%28MAP-Elites%29+%2F+Quality+Diversity%3A+A+New+Frontier+for+Evolutionary+Computation 16. Reaching Beyond the Mode: RL for Distributional Reasoning in Language Models — Isha Puri, Mehul Damani, Idan Shenfeld, Marzyeh Ghassemi, Jacob Andreas, Yoon Kim, 2026 https://scholar.google.com/scholar?q=Reaching+Beyond+the+Mode%3A+RL+for+Distributional+Reasoning+in+Language+Models 17. KL-Regularized Reinforcement Learning is Designed to Mode Collapse — Anthony GX-Chen, Jatin Prakash, Jeff Guo, Rob Fergus, Rajesh Ranganath, 2025 https://scholar.google.com/scholar?q=KL-Regularized+Reinforcement+Learning+is+Designed+to+Mode+Collapse 18. The Best of N Worlds: Aligning Reinforcement Learning with Best-of-N Sampling via max@k Optimisation — Farid Bagirov, Mikhail Arkhipov, Ksenia Sycheva, Evgeniy Glukhov, Egor Bogomolov, 2025 https://scholar.google.com/scholar?q=The+Best+of+N+Worlds%3A+Aligning+Reinforcement+Learning+with+Best-of-N+Sampling+via+max%40k+Optimisation 19. Maximum Likelihood Reinforcement Learning (MaxRL) — Fahim Tajwar, Guanning Zeng, Yueer Zhou, Yuda Song, et al., 2026 https://scholar.google.com/scholar?q=Maximum+Likelihood+Reinforcement+Learning+%28MaxRL%29 20. GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization — Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, et al. (NVIDIA), 2026 https://scholar.google.com/scholar?q=GDPO%3A+Group+reward-Decoupled+Normalization+Policy+Optimization+for+Multi-reward+RL+Optimization 21. A Survey of Multi-Objective Sequential Decision-Making / A Practical Guide to Multi-Objective RL and Planning — Roijers, Vamplew, Whiteson, Dazeley (2013); Hayes et al. (2022), 2013/2022 https://scholar.google.com/scholar?q=A+Survey+of+Multi-Objective+Sequential+Decision-Making+%2F+A+Practical+Guide+to+Multi-Objective+RL+and+Planning 22. Lexicase selection (Spector 2012; La Cava et al. 2019; Ni et al. 2024, DALex) — Lee Spector; William La Cava, Thomas Helmuth, et al.; Andrew Ni, Li Ding, Lee Spector, 2012/2019/2024 https://scholar.google.com/scholar?q=Lexicase+selection+%28Spector+2012%3B+La+Cava+et+al.+2019%3B+Ni+et+al.+2024%2C+DALex%29 23. Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models / BOND / BoNBoN / InfAlign — Chow et al.; Sessa et al.; Gui, Gârbacea, Veitch; Balashankar et al., 2024-2025 https://scholar.google.com/scholar?q=Inference-Aware+Fine-Tuning+for+Best-of-N+Sampling+in+Large+Language+Models+%2F+BOND+%2F+BoNBoN+%2F+InfAlign 24. AlphaEvolve / FunSearch / ShinkaEvolve / OpenEvolve — Novikov et al.; Romera-Paredes et al.; Lange, Imajuku, Cetin; Sharma, 2023-2025 https://scholar.google.com/scholar?q=AlphaEvolve+%2F+FunSearch+%2F+ShinkaEvolve+%2F+OpenEvolve 25. Rewards-in-Context: Multi-objective Alignment of Foundation Models with Dynamic Preference Adjustment — Rui Yang, Xiaoman Pan, Feng Luo, et al., 2024 https://scholar.google.com/scholar?q=Rewards-in-Context%3A+Multi-objective+Alignment+of+Foundation+Models+with+Dynamic+Preference+Adjustment 26. Rewarded Soups / Directional Preference Alignment — Rame et al. (2023); Wang et al. (2024), 2023/2024 https://scholar.google.com/scholar?q=Rewarded+Soups+%2F+Directional+Preference+Alignment 27. DARLING (diversity-aware RL) and entropy-preserving GRPO variants (DAPO clip-higher) — Li et al.; Yu et al., 2025 https://scholar.google.com/scholar?q=DARLING+%28diversity-aware+RL%29+and+entropy-preserving+GRPO+variants+%28DAPO+clip-higher%29 Interactive Visualization: Vector Policy Optimization: Training for Diversity to Improve Test-Time Search

    Transcription

22. Harness2: Recursive Self-Improving Agent Harnesses Without Verifiers

    31 octobre 2026À suivre

    This episode examines "Harness 2," a paper on recursive agent harnessing. Its premise is that the wrapper around a frozen model can be improved automatically on a stream of unlabeled tasks, with no ground-truth …

    Notes de l’épisode

    This episode examines "Harness 2," a paper on recursive agent harnessing. Its premise is that the wrapper around a frozen model can be improved automatically on a stream of unlabeled tasks, with no ground-truth verifier. The "harness" here is eight editable files: prompts, guardrails, memory, skills, tools, scripts, subagents, and middleware. The discussion places the work against prior approaches. These include SWE-agent's interface findings, Meta-Harness's freeze-after-search approach, per-instance scaling, ReasoningBank, Reflexion, and the Darwin Gödel Machine. It then explains the two coupled loops. Task-level refinement edits the execution harness for the current task. Harnessing-level recursion folds the lessons into a separate improvement harness, made of a library of adopted harnesses plus an EXPERIENCE.md file, while each new task still starts from a common base. The episode begins a walkthrough of the Propose, Probe, Compose refinement step, using a JobBench case that rose from 43.1 to 77.8. Listeners interested in self-improving agents will find it useful for seeing why self-correction without external feedback is hard and how this design tries to work around that. Sources: 1. Harness^2: Recursive Agent Harnessing for an Open World — Ruiyao Xu, Yanfei Chen, Zhongying CuiZhu, Bhavana Dalvi Mishra, Yifei Ming, Han Yu, Rujun Han, Chen-Yu Lee, Tomas Pfister, 2026 http://arxiv.org/abs/2610.14224 2. Reflexion: Language Agents with Verbal Reinforcement Learning — Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao, 2023 https://scholar.google.com/scholar?q=Reflexion%3A+Language+Agents+with+Verbal+Reinforcement+Learning 3. Automated Design of Agentic Systems (ADAS) — Shengran Hu, Cong Lu, Jeff Clune, 2024 https://scholar.google.com/scholar?q=Automated+Design+of+Agentic+Systems+%28ADAS%29 4. Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents — Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, Jeff Clune, 2025 https://scholar.google.com/scholar?q=Darwin+G%C3%B6del+Machine%3A+Open-Ended+Evolution+of+Self-Improving+Agents 5. ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory — Siru Ouyang et al. (Google Cloud AI Research and collaborators), 2025 https://scholar.google.com/scholar?q=ReasoningBank%3A+Scaling+Agent+Self-Evolving+with+Reasoning+Memory 6. Self-Refine: Iterative Refinement with Self-Feedback — Aman Madaan, Niket Tandon, Peter Clark, Yiming Yang, et al., 2023 https://scholar.google.com/scholar?q=Self-Refine%3A+Iterative+Refinement+with+Self-Feedback 7. Large Language Models Cannot Self-Correct Reasoning Yet — Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, Denny Zhou, 2023 https://scholar.google.com/scholar?q=Large+Language+Models+Cannot+Self-Correct+Reasoning+Yet 8. Large Language Models as Optimizers (OPRO) — Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, Xinyun Chen, 2023 https://scholar.google.com/scholar?q=Large+Language+Models+as+Optimizers+%28OPRO%29 9. Meta-Harness (Lee et al.) — Lee et al. (as cited in the Harness2 paper; full author list not verified), 2026 https://scholar.google.com/scholar?q=Meta-Harness+%28Lee+et+al.%29 10. SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering — John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, Ofir Press, 2024 https://scholar.google.com/scholar?q=SWE-agent%3A+Agent-Computer+Interfaces+Enable+Automated+Software+Engineering 11. Meta-Harness: End-to-end optimization of model harnesses — Y. Lee, R. S. Nair, Q. Zhang, K. Lee, O. Khattab, C. Finn, 2026 https://scholar.google.com/scholar?q=Meta-Harness%3A+End-to-end+optimization+of+model+harnesses 12. Rethinking the evaluation of harness evolution for agents — Y. Wang, H. Zhu, Z. Hu, et al., 2026 https://scholar.google.com/scholar?q=Rethinking+the+evaluation+of+harness+evolution+for+agents 13. Recursive harness self-improvement; and TTHE: Test-time harness evolution — H. Lee, J. Xu, J. Seely, D. Lee, M. Zaharia, Y. Tang (2026a); J. Nie et al. (2026b), 2026 https://scholar.google.com/scholar?q=Recursive+harness+self-improvement%3B+and+TTHE%3A+Test-time+harness+evolution 14. Hyperagents; Darwin Gödel Machine; Self-taught Optimizer (STOP) — J. Zhang et al. (Hyperagents 2026, DGM 2026); E. Zelikman et al. (STOP 2024), 2024-2026 https://scholar.google.com/scholar?q=Hyperagents%3B+Darwin+G%C3%B6del+Machine%3B+Self-taught+Optimizer+%28STOP%29 15. Harness continual learning (HCL) and Hierarchical self-improvement (HSI) — B. Kang et al. (2026); T. Zhou (2026), 2026 https://scholar.google.com/scholar?q=Harness+continual+learning+%28HCL%29+and+Hierarchical+self-improvement+%28HSI%29 16. GEPA: Reflective prompt evolution can outperform reinforcement learning — L. A. Agrawal et al., 2026 https://scholar.google.com/scholar?q=GEPA%3A+Reflective+prompt+evolution+can+outperform+reinforcement+learning 17. Recursive Self-Aggregation Unlocks Deep Thinking in LLMs — S. Venkatraman et al., 2025 https://scholar.google.com/scholar?q=Recursive+Self-Aggregation+Unlocks+Deep+Thinking+in+LLMs 18. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena; LLM Evaluators Recognize and Favor Their Own Generations — L. Zheng et al. (2023); A. Panickssery et al. (2024), 2023-2024 https://scholar.google.com/scholar?q=Judging+LLM-as-a-Judge+with+MT-Bench+and+Chatbot+Arena%3B+LLM+Evaluators+Recognize+and+Favor+Their+Own+Generations 19. Scaling Laws for Reward Model Overoptimization — L. Gao, J. Schulman, J. Hilton, 2023 https://scholar.google.com/scholar?q=Scaling+Laws+for+Reward+Model+Overoptimization 20. Reflexion: Language agents with verbal reinforcement learning; Self-Refine — N. Shinn et al. (2023); A. Madaan et al. (2023), 2023 https://scholar.google.com/scholar?q=Reflexion%3A+Language+agents+with+verbal+reinforcement+learning%3B+Self-Refine 21. Automated Design of Agentic Systems (ADAS); AFlow; DSPy — S. Hu, C. Lu, J. Clune (2025); J. Zhang et al. (2025); O. Khattab et al. (2024), 2024-2025 https://scholar.google.com/scholar?q=Automated+Design+of+Agentic+Systems+%28ADAS%29%3B+AFlow%3B+DSPy Interactive Visualization: Harness2: Recursive Self-Improving Agent Harnesses Without Verifiers

    Transcription

23. Sliding-Window Attention Beats Linear Attention Without Any Training

    31 octobre 2026À suivre

    This episode examines a paper claiming that a pretrained LLM can be cut down to a sliding window of the last 64 tokens plus four "attention sink" starting tokens, with no training, and still keep about 99% of …

    Notes de l’épisode

    This episode examines a paper claiming that a pretrained LLM can be cut down to a sliding window of the last 64 tokens plus four "attention sink" starting tokens, with no training, and still keep about 99% of its benchmark score. That result would match months of work on retrofitting models into linear attention. The discussion builds up the background: the KV cache and how eviction differs from fixed-size linear-attention state, why StreamingLLM's sinks keep windowing from collapsing, and how retrofit costs range from roughly 20 million to 100 billion training tokens. It then questions what "99% recovery" means, noting that log-likelihood multiple-choice benchmarks like MMLU and HellaSwag mostly need only nearby context, unlike generative tasks such as GSM8K and HumanEval. It also asks whether the comparison is fair when memory budgets aren't held equal. Listeners get a skeptical look at whether a simple free baseline is outperforming retrofit methods, or whether the benchmarks are just too local to tell. Sources: 1. Sliding-window beats linear attention — Alexia Jolicoeur-Martineau, Rhea Sanjay Sukthanker, Pashmina Cameron, Emy Gervais, 2026 http://arxiv.org/abs/2608.28444 2. Efficient Streaming Language Models with Attention Sinks — Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, Mike Lewis, 2023 (ICLR 2024) https://scholar.google.com/scholar?q=Efficient+Streaming+Language+Models+with+Attention+Sinks 3. Longformer: The Long-Document Transformer — Iz Beltagy, Matthew E. Peters, Arman Cohan, 2020 https://scholar.google.com/scholar?q=Longformer%3A+The+Long-Document+Transformer 4. Mistral 7B — Albert Q. Jiang et al. (Mistral AI), 2023 https://scholar.google.com/scholar?q=Mistral+7B 5. Gemma 2: Improving Open Language Models at a Practical Size (and Gemma 3 Technical Report) — Gemma Team, Google DeepMind, 2024 / 2025 https://scholar.google.com/scholar?q=Gemma+2%3A+Improving+Open+Language+Models+at+a+Practical+Size+%28and+Gemma+3+Technical+Report%29 6. Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention — Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, François Fleuret, 2020 https://scholar.google.com/scholar?q=Transformers+are+RNNs%3A+Fast+Autoregressive+Transformers+with+Linear+Attention 7. Finetuning Pretrained Transformers into RNNs — Jungo Kasai, Hao Peng, Yizhe Zhang, Dani Yogatama, Gabriel Ilharco, Nikolaos Pappas, Yi Mao, Weizhu Chen, Noah A. Smith, 2021 https://scholar.google.com/scholar?q=Finetuning+Pretrained+Transformers+into+RNNs 8. LoLCATs: On Low-Rank Linearizing of Large Language Models — Michael Zhang, Simran Arora, Rahul Chalamala, Alan Wu, Benjamin Spector, Aaryan Singhal, Krithik Ramesh, Christopher Ré, 2024 (ICLR 2025) https://scholar.google.com/scholar?q=LoLCATs%3A+On+Low-Rank+Linearizing+of+Large+Language+Models 9. Transformers to SSMs: Distilling Quadratic Knowledge to Subquadratic Models (MOHAWK) — Aviv Bick, Kevin Y. Li, Eric P. Xing, J. Zico Kolter, Albert Gu, 2024 https://scholar.google.com/scholar?q=Transformers+to+SSMs%3A+Distilling+Quadratic+Knowledge+to+Subquadratic+Models+%28MOHAWK%29 10. The Mamba in the Llama: Distilling and Accelerating Hybrid Models — Junxiong Wang, Daniele Paliotta, Avner May, Alexander M. Rush, Tri Dao, 2024 https://scholar.google.com/scholar?q=The+Mamba+in+the+Llama%3A+Distilling+and+Accelerating+Hybrid+Models 11. Efficient Streaming Language Models with Attention Sinks (StreamingLLM) — Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, Mike Lewis, 2024 https://scholar.google.com/scholar?q=Efficient+Streaming+Language+Models+with+Attention+Sinks+%28StreamingLLM%29 12. Why do LLMs attend to the first token? — Federico Barbero, Alvaro Arroyo, Xiangming Gu, Christos Perivolaropoulos, Michael Bronstein, Petar Veličković, Razvan Pascanu, 2025 https://scholar.google.com/scholar?q=Why+do+LLMs+attend+to+the+first+token%3F 13. Short window attention enables long-term memorization — Loïc Cabannes, Maximilian Beck, Maria Lomeli, Gergely Szilvasy, Matthijs Douze, Jade Copet, Pierre-Emmanuel Mazaré, Gabriel Synnaeve, Hervé Jégou, 2026 https://scholar.google.com/scholar?q=Short+window+attention+enables+long-term+memorization 14. SWAA: Sliding Window Attention Adaptation for Efficient and Quality Preserving Long Context Processing — Yijiong Yu, Jiale Liu, Qingyun Wu, Huazheng Wang, Ji Pei, 2025 https://scholar.google.com/scholar?q=SWAA%3A+Sliding+Window+Attention+Adaptation+for+Efficient+and+Quality+Preserving+Long+Context+Processing 15. STILL: Selecting Tokens for Intra-Layer Hybrid Attention to Linearize LLMs — Weikang Meng, Liangyu Huo, Yadan Luo, Jiawen Guan, Jingyi Zhang, Yingjian Li, Zheng Zhang, 2026 https://scholar.google.com/scholar?q=STILL%3A+Selecting+Tokens+for+Intra-Layer+Hybrid+Attention+to+Linearize+LLMs 16. Transformers to SSMs: Distilling Quadratic Knowledge to Subquadratic Models (MOHAWK) and Llamba: Scaling Distilled Recurrent Models — Aviv Bick, Kevin Y. Li, Eric P. Xing, J. Zico Kolter, Albert Gu; Aviv Bick, Tobias Katsch, Nimit Sohoni, Arjun Desai, Albert Gu, 2024/2025 https://scholar.google.com/scholar?q=Transformers+to+SSMs%3A+Distilling+Quadratic+Knowledge+to+Subquadratic+Models+%28MOHAWK%29+and+Llamba%3A+Scaling+Distilled+Recurrent+Models 17. Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models — Soham De, Samuel L. Smith, Anushan Fernando, et al., 2024 https://scholar.google.com/scholar?q=Griffin%3A+Mixing+Gated+Linear+Recurrences+with+Local+Attention+for+Efficient+Language+Models 18. H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models; SnapKV: LLM Knows What You Are Looking For Before Generation — Zhenyu Zhang et al.; Yuhong Li et al., 2023/2024 https://scholar.google.com/scholar?q=H2O%3A+Heavy-Hitter+Oracle+for+Efficient+Generative+Inference+of+Large+Language+Models%3B+SnapKV%3A+LLM+Knows+What+You+Are+Looking+For+Before+Generation 19. RULER: What's the Real Context Size of Your Long-Context Language Models?; BABILong — Cheng-Ping Hsieh et al.; Yuri Kuratov et al., 2024 https://scholar.google.com/scholar?q=RULER%3A+What%27s+the+Real+Context+Size+of+Your+Long-Context+Language+Models%3F%3B+BABILong 20. Gated Delta Networks: Improving Mamba2 with Delta Rule; Parallelizing Linear Transformers with the Delta Rule — Songlin Yang, Jan Kautz, Ali Hatamizadeh; Songlin Yang et al., 2025/2024 https://scholar.google.com/scholar?q=Gated+Delta+Networks%3A+Improving+Mamba2+with+Delta+Rule%3B+Parallelizing+Linear+Transformers+with+the+Delta+Rule Interactive Visualization: Sliding-Window Attention Beats Linear Attention Without Any Training

    Transcription

24. The Illusion of State in State-Space Models

    30 octobre 2026À suivre

    This episode examines "The Illusion of State in State-Space Models," which asks whether S4 and Mamba, despite their recurrent form, can track state any better than transformers. It frames state tracking as the …

    Notes de l’épisode

    This episode examines "The Illusion of State in State-Space Models," which asks whether S4 and Mamba, despite their recurrent form, can track state any better than transformers. It frames state tracking as the monoid word problem, from a single parity switch up to the non-commutative group S5 (five people swapping coins). Barrington's theorem makes S5 complete for NC1, while constant-depth transformers and SSMs both land in TC0. The argument rests on the unproven conjecture that TC0 ≠ NC1, and it assumes log-precision numbers and worst-case inputs. The discussion explains why linear, nonlinearity-free recurrences can be computed with a parallel scan, and why that same property caps their expressive power. It then walks through the proof that S4's fixed matrix powers and Mamba's input-dependent diagonal matrices both stay within TC0. Listeners get a clear account of why "has a recurrent form" does not imply "has recurrent power." Sources: 1. The Illusion of State in State-Space Models — William Merrill, Jackson Petty, Ashish Sabharwal, 2024 http://arxiv.org/abs/2404.08819 2. Transformers Learn Shortcuts to Automata — Bingbin Liu, Jordan T. Ash, Surbhi Goel, Akshay Krishnamurthy, Cyril Zhang, 2023 https://scholar.google.com/scholar?q=Transformers+Learn+Shortcuts+to+Automata 3. Bounded-width polynomial-size branching programs recognize exactly those languages in NC1 — David A. Barrington, 1986 (conference 1985) https://scholar.google.com/scholar?q=Bounded-width+polynomial-size+branching+programs+recognize+exactly+those+languages+in+NC1 4. The Parallelism Tradeoff: Limitations of Log-Precision Transformers — William Merrill, Ashish Sabharwal, 2023 https://scholar.google.com/scholar?q=The+Parallelism+Tradeoff%3A+Limitations+of+Log-Precision+Transformers 5. The Illusion of State in State-Space Models — William Merrill, Jackson Petty, Ashish Sabharwal, 2024 https://scholar.google.com/scholar?q=The+Illusion+of+State+in+State-Space+Models 6. The Expressive Power of Transformers with Chain of Thought — William Merrill, Ashish Sabharwal, 2024 https://scholar.google.com/scholar?q=The+Expressive+Power+of+Transformers+with+Chain+of+Thought 7. Mamba: Linear-Time Sequence Modeling with Selective State Spaces — Albert Gu, Tri Dao, 2023 https://scholar.google.com/scholar?q=Mamba%3A+Linear-Time+Sequence+Modeling+with+Selective+State+Spaces 8. Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality (Mamba-2) — Tri Dao, Albert Gu, 2024 https://scholar.google.com/scholar?q=Transformers+are+SSMs%3A+Generalized+Models+and+Efficient+Algorithms+Through+Structured+State+Space+Duality+%28Mamba-2%29 9. Unlocking State-Tracking in Linear RNNs Through Negative Eigenvalues — Riccardo Grazzi, Julien Siems, Arber Zela, Jörg K.H. Franke, Frank Hutter, Massimiliano Pontil, 2024 https://scholar.google.com/scholar?q=Unlocking+State-Tracking+in+Linear+RNNs+Through+Negative+Eigenvalues 10. Combining Recurrent, Convolutional, and Continuous-time Models with Linear State Space Layers (LSSL) — Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, Christopher Ré, 2021 https://scholar.google.com/scholar?q=Combining+Recurrent%2C+Convolutional%2C+and+Continuous-time+Models+with+Linear+State+Space+Layers+%28LSSL%29 11. Liquid Structural State-Space Models — Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, Daniela Rus, 2023 https://scholar.google.com/scholar?q=Liquid+Structural+State-Space+Models 12. Chess as a Testbed for Language Model State Tracking — Shubham Toshniwal, Sam Wiseman, Karen Livescu, Kevin Gimpel, 2021 https://scholar.google.com/scholar?q=Chess+as+a+Testbed+for+Language+Model+State+Tracking 13. Entity Tracking in Language Models — Najoung Kim, Sebastian Schuster, 2023 https://scholar.google.com/scholar?q=Entity+Tracking+in+Language+Models 14. Parallelizing Linear Transformers with the Delta Rule over Sequence Length (DeltaNet) — Songlin Yang, Bailin Wang, Yu Zhang, Yikang Shen, Yoon Kim, 2024 https://scholar.google.com/scholar?q=Parallelizing+Linear+Transformers+with+the+Delta+Rule+over+Sequence+Length+%28DeltaNet%29 15. The Expressive Capacity of State Space Models: A Formal Language Perspective — Yash Sarrof, Yana Veitsman, Michael Hahn, 2024 https://scholar.google.com/scholar?q=The+Expressive+Capacity+of+State+Space+Models%3A+A+Formal+Language+Perspective 16. The Illusion of Thought / Transformers, parallel computation, and logarithmic depth — Clayton Sanford, Daniel Hsu, Matus Telgarsky, 2024 https://scholar.google.com/scholar?q=The+Illusion+of+Thought+%2F+Transformers%2C+parallel+computation%2C+and+logarithmic+depth 17. Barrington's theorem: Bounded-width polynomial-size branching programs recognize exactly those languages in NC1 — David A. Barrington, 1989 https://scholar.google.com/scholar?q=Barrington%27s+theorem%3A+Bounded-width+polynomial-size+branching+programs+recognize+exactly+those+languages+in+NC1 Interactive Visualization: The Illusion of State in State-Space Models

    Transcription

25. LLM Evolutionary Search Finds Cheaper Quantum Arithmetic Circuits

    30 octobre 2026À suivre

    This episode examines a paper that uses an LLM-driven evolutionary search (OpenEvolve with Claude Mythos 5) to find quantum arithmetic circuits needing fewer Toffoli gates and fewer qubits than the best …

    Notes de l’épisode

    This episode examines a paper that uses an LLM-driven evolutionary search (OpenEvolve with Claude Mythos 5) to find quantum arithmetic circuits needing fewer Toffoli gates and fewer qubits than the best published designs across eight tasks. These tasks include decoders for decoded quantum interferometry, modular inversion, and chemistry and lattice-gauge workloads. It explains why quantum arithmetic is hard to write: coherent control flow forces every branch to run, reversibility requires uncomputation, and a stated failure budget can be traded for savings. It also covers how candidates are checked classically by tracking basis states and phases, so hallucinated circuits are rejected by an exact evaluator. The discussion places the work against FunSearch, AlphaEvolve and rewrite-based optimizers, and notes the trade-off that correctness is statistical rather than proved. It also flags the Anthropic affiliation and the simplification of equally weighting gate count and qubit count into one cost metric. Sources: 1. Discovering state-of-the-art quantum circuits with large language models — Andi Gu, Alek H. Kemeny, Susanne F. Yelin, Mikhail D. Lukin, 2026 http://arxiv.org/abs/2610.22367 2. How to factor 2048 bit RSA integers in 8 hours using 20 million noisy qubits — Craig Gidney, Martin Ekera, 2019 (Quantum 2021) https://scholar.google.com/scholar?q=How+to+factor+2048+bit+RSA+integers+in+8+hours+using+20+million+noisy+qubits 3. Encoding Electronic Spectra in Quantum Circuits with Linear T Complexity — Ryan Babbush, Craig Gidney, Dominic Berry, Nathan Wiebe, Jarrod McClean, Alexandru Paler, Austin Fowler, Hartmut Neven, 2018 https://scholar.google.com/scholar?q=Encoding+Electronic+Spectra+in+Quantum+Circuits+with+Linear+T+Complexity 4. AlphaEvolve: A coding agent for scientific and algorithmic discovery (and the preceding FunSearch: Mathematical discoveries from program search with large language models) — Alexander Novikov, Ngan Vu, Marvin Eisenberger, et al. (Google DeepMind); FunSearch: Bernardino Romera-Paredes, Mohammadamin Barekatain, et al., 2025 (AlphaEvolve); 2023 (FunSearch, Nature 2024) https://scholar.google.com/scholar?q=AlphaEvolve%3A+A+coding+agent+for+scientific+and+algorithmic+discovery+%28and+the+preceding+FunSearch%3A+Mathematical+discoveries+from+program+search+with+large+language+models%29 5. Discovering faster matrix multiplication algorithms with reinforcement learning (AlphaTensor) and Quantum Circuit Optimization with AlphaTensor (AlphaTensor-Quantum) — Alhussein Fawzi et al.; Francisco J. R. Ruiz et al. (DeepMind), 2022 (AlphaTensor); 2024 (AlphaTensor-Quantum, Nature Machine Intelligence 2025) https://scholar.google.com/scholar?q=Discovering+faster+matrix+multiplication+algorithms+with+reinforcement+learning+%28AlphaTensor%29+and+Quantum+Circuit+Optimization+with+AlphaTensor+%28AlphaTensor-Quantum%29 6. Logical Reversibility of Computation — Charles H. Bennett, 1973 https://scholar.google.com/scholar?q=Logical+Reversibility+of+Computation 7. Quantum Resource Estimates for Computing Elliptic Curve Discrete Logarithms — Martin Roetteler, Michael Naehrig, Krysta M. Svore, Kristin Lauter, 2017 https://scholar.google.com/scholar?q=Quantum+Resource+Estimates+for+Computing+Elliptic+Curve+Discrete+Logarithms 8. Fast constant-time gcd computation and modular inversion — Daniel J. Bernstein, Bo-Yin Yang, 2019 https://scholar.google.com/scholar?q=Fast+constant-time+gcd+computation+and+modular+inversion 9. Reversible Residual Network: Backpropagation Without Storing Activations (RevNet) — Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse, 2017 https://scholar.google.com/scholar?q=Reversible+Residual+Network%3A+Backpropagation+Without+Storing+Activations+%28RevNet%29 10. Mathematical discoveries from program search with large language models (FunSearch) — Romera-Paredes et al., 2023/2024 https://scholar.google.com/scholar?q=Mathematical+discoveries+from+program+search+with+large+language+models+%28FunSearch%29 11. AlphaEvolve: A coding agent for scientific and algorithmic discovery — Novikov et al. (Google DeepMind), 2025 https://scholar.google.com/scholar?q=AlphaEvolve%3A+A+coding+agent+for+scientific+and+algorithmic+discovery 12. Quartz / PyZX / TODD / T-par style circuit optimizers — Xu et al. (Quartz); Kissinger & van de Wetering (PyZX); Heyfron & Campbell (TODD); Amy, Maslov & Mosca (T-par), 2019-2022 https://scholar.google.com/scholar?q=Quartz+%2F+PyZX+%2F+TODD+%2F+T-par+style+circuit+optimizers 13. Faster quantum arithmetic / AlphaTensor-Quantum: Quantum circuit optimization with deep reinforcement learning — Ruiz et al. (DeepMind), 2024/2025 https://scholar.google.com/scholar?q=Faster+quantum+arithmetic+%2F+AlphaTensor-Quantum%3A+Quantum+circuit+optimization+with+deep+reinforcement+learning 14. Decoded Quantum Interferometry — Jordan et al., 2024 https://scholar.google.com/scholar?q=Decoded+Quantum+Interferometry 15. Compiling decoded quantum interferometry / reversible Reed-Solomon decoders for DQI (Khattar et al.) — Khattar et al., 2025 https://scholar.google.com/scholar?q=Compiling+decoded+quantum+interferometry+%2F+reversible+Reed-Solomon+decoders+for+DQI+%28Khattar+et+al.%29 16. Fast constant-time gcd computation and modular inversion (safegcd / divsteps) — Bernstein & Yang, 2019 https://scholar.google.com/scholar?q=Fast+constant-time+gcd+computation+and+modular+inversion+%28safegcd+%2F+divsteps%29 17. How to factor 2048 bit RSA integers with less than a million noisy qubits — Gidney, 2025 https://scholar.google.com/scholar?q=How+to+factor+2048+bit+RSA+integers+with+less+than+a+million+noisy+qubits 18. Halving the cost of quantum addition / Measurement-based uncomputation of quantum circuits — Gidney (2018); Berry et al. (2019), 2018/2019 https://scholar.google.com/scholar?q=Halving+the+cost+of+quantum+addition+%2F+Measurement-based+uncomputation+of+quantum+circuits 19. Coset representation of modular integers / approximate residue arithmetic — Zalka; Gidney, 2006/2019 https://scholar.google.com/scholar?q=Coset+representation+of+modular+integers+%2F+approximate+residue+arithmetic 20. Efficient implementations of reversible computation / Bennett pebble game space-time trade-offs — Bennett; Knill; Chan & Gidney, 1989/1995/2013 https://scholar.google.com/scholar?q=Efficient+implementations+of+reversible+computation+%2F+Bennett+pebble+game+space-time+trade-offs 21. The Reversible Residual Network: Backpropagation Without Storing Activations — Gomez et al., 2017 https://scholar.google.com/scholar?q=The+Reversible+Residual+Network%3A+Backpropagation+Without+Storing+Activations Interactive Visualization: LLM Evolutionary Search Finds Cheaper Quantum Arithmetic Circuits

    Transcription

26. FastPersist: Making Per-Iteration Deep Learning Checkpointing Nearly Free

    30 octobre 2026À suivre

    This episode examines FastPersist, a DeepSpeed paper on making deep learning model checkpointing fast enough to save every training iteration. It lays out why checkpoint time has grown to 50–96% of iteration …

    Notes de l’épisode

    This episode examines FastPersist, a DeepSpeed paper on making deep learning model checkpointing fast enough to save every training iteration. It lays out why checkpoint time has grown to 50–96% of iteration time as data parallelism scaled, and why a checkpoint is so large: mixed-precision Adam state works out to about 14 bytes per parameter, or 173 GB for a 13B model. The discussion traces the slowness to torch.save issuing small, unaligned writes. That leaves NVMe drives well below peak, with the baseline reaching under 20% of the 24.8 GB/s a RAID-0 array can deliver. It also covers the structural opportunity to overlap checkpointing with the next forward and backward passes, and to spread writes across data-parallel replicas that hold identical state. The episode compares the approach with CheckFreq, Check-N-Run and Gemini, and it questions whether checkpointing every iteration is the right target. It also notes what the paper leaves out, such as SSD write-cache behavior and flash wear. Sources: 1. FastPersist: Accelerating Model Checkpointing in Deep Learning — Guanhua Wang, Olatunji Ruwase, Bing Xie, Yuxiong He, 2024 http://arxiv.org/abs/2406.13768 2. CheckFreq: Frequent, Fine-Grained DNN Checkpointing — Jayashree Mohan, Amar Phanishayee, Vijay Chidambaram, 2021 https://scholar.google.com/scholar?q=CheckFreq%3A+Frequent%2C+Fine-Grained+DNN+Checkpointing 3. Gemini: Fast Failure Recovery in Distributed Training with In-Memory Checkpoints — Zhuang Wang, Zhen Jia, Shuai Zheng, Zhen Zhang, Xinwei Fu, T. S. Eugene Ng, Yida Wang, 2023 https://scholar.google.com/scholar?q=Gemini%3A+Fast+Failure+Recovery+in+Distributed+Training+with+In-Memory+Checkpoints 4. Check-N-Run: A Checkpointing System for Training Deep Learning Recommendation Models — Assaf Eisenman, Kiran Kumar Matam, Steven Ingram, Dheevatsa Mudigere, Raghuraman Krishnamoorthi, Krishnakumar Nair, Misha Smelyanskiy, Murali Annavaram, 2022 https://scholar.google.com/scholar?q=Check-N-Run%3A+A+Checkpointing+System+for+Training+Deep+Learning+Recommendation+Models 5. ByteCheckpoint: A Unified Checkpointing System for Large Foundation Model Development — Borui Wan, Mingji Han, Yiyao Sheng, Yanghua Peng, Haibin Lin, Mofan Zhang, Zhichao Lai, Menghan Yu, Junda Zhang, Zuquan Song, Xin Liu, Chuan Wu, 2024 https://scholar.google.com/scholar?q=ByteCheckpoint%3A+A+Unified+Checkpointing+System+for+Large+Foundation+Model+Development 6. ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning — Samyam Rajbhandari, Olatunji Ruwase, Jeff Rasley, Shaden Smith, Yuxiong He, 2021 https://scholar.google.com/scholar?q=ZeRO-Infinity%3A+Breaking+the+GPU+Memory+Wall+for+Extreme+Scale+Deep+Learning 7. Characterizing Deep Learning Training Workloads on Alibaba-PAI / Understanding Data Storage and Ingestion for Large-Scale Deep Recommendation Model Training — Mark Zhao, Niket Agarwal, Aarti Basant, Buğra Gedik, Satadru Pan, Mustafa Ozdal, Rakesh Komuravelli, Jerry Pan, Tianshu Bao, Haowei Lu, Sundaram Narayanan, Jack Langman, Kevin Wilfong, Harsha Rastogi, Carole-Jean Wu, Christos Kozyrakis, Parik Pol, 2022 https://scholar.google.com/scholar?q=Characterizing+Deep+Learning+Training+Workloads+on+Alibaba-PAI+%2F+Understanding+Data+Storage+and+Ingestion+for+Large-Scale+Deep+Recommendation+Model+Training 8. What Modern NVMe Storage Can Do, And How To Exploit It: High-Performance I/O for High-Performance Storage Engines — Gabriel Haas, Viktor Leis, 2023 https://scholar.google.com/scholar?q=What+Modern+NVMe+Storage+Can+Do%2C+And+How+To+Exploit+It%3A+High-Performance+I%2FO+for+High-Performance+Storage+Engines 9. PyTorch Distributed Checkpoint (DCP) and async_save — PyTorch team (Meta), 2023-2024 https://scholar.google.com/scholar?q=PyTorch+Distributed+Checkpoint+%28DCP%29+and+async_save 10. Reducing Activation Recomputation / OPT-175B logbook and Llama 3 herd paper (failure analyses) — Susan Zhang et al.; Llama Team, Meta AI, 2022; 2024 https://scholar.google.com/scholar?q=Reducing+Activation+Recomputation+%2F+OPT-175B+logbook+and+Llama+3+herd+paper+%28failure+analyses%29 11. The Young/Daly optimal checkpoint interval and multi-level checkpointing (e.g., SCR, FTI) — J. Young; J. Daly; Moody et al.; Bautista-Gomez et al., 1974; 2006; 2010-2011 https://scholar.google.com/scholar?q=The+Young%2FDaly+optimal+checkpoint+interval+and+multi-level+checkpointing+%28e.g.%2C+SCR%2C+FTI%29 12. GPUDirect Storage / cuFile and SPIN — NVIDIA; Bergman, Brokhman, Cohen, Silberstein, 2017-2022 https://scholar.google.com/scholar?q=GPUDirect+Storage+%2F+cuFile+and+SPIN 13. ReCycle / Oobleck / Bamboo style resilient training without checkpoint restart — Various, 2023-2024 https://scholar.google.com/scholar?q=ReCycle+%2F+Oobleck+%2F+Bamboo+style+resilient+training+without+checkpoint+restart Interactive Visualization: FastPersist: Making Per-Iteration Deep Learning Checkpointing Nearly Free

    Transcription

27. Latent Compaction with Test-Time Training: Learning Compact KV Caches

    30 octobre 2026À suivre

    This episode examines "Latent Compaction with Test-Time Training," a preprint that asks whether a frozen model can shrink a document's KV cache by 16 to 32 times by running gradient descent on a small set of …

    Notes de l’épisode

    This episode examines "Latent Compaction with Test-Time Training," a preprint that asks whether a frozen model can shrink a document's KV cache by 16 to 32 times by running gradient descent on a small set of new key-value entries. The compaction happens before any questions are known. The discussion explains why the cache bottlenecks long-context memory and batching, and why text summaries lose details like file paths or IDs. It then contrasts the approach with eviction methods such as H2O, SnapKV, PyramidKV and KVZip, which can only select existing entries and never create blended ones. It also places the work among prefix-tuning, Cartridges, Attention Matching and other test-time training approaches, and clarifies that the "32x" figure counts cache slots, not bytes, with no measured speedup or memory savings. Listeners get a clear primer on query-aware versus query-agnostic compression, along with a skeptical look at what the headline RULER result does and doesn't show. Sources: 1. Latent Compaction with Test-Time Training — Yingfa Chen, Xingyu Shen, Chaojun Xiao, Xu Han, 2026 http://arxiv.org/abs/2610.20582v1 2. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts — Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, Moritz Hardt, 2020 https://scholar.google.com/scholar?q=Test-Time+Training+with+Self-Supervision+for+Generalization+under+Distribution+Shifts 3. Learning to (Learn at Test Time): RNNs with Expressive Hidden States — Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, Carlos Guestrin, 2024 https://scholar.google.com/scholar?q=Learning+to+%28Learn+at+Test+Time%29%3A+RNNs+with+Expressive+Hidden+States 4. The Surprising Effectiveness of Test-Time Training for Few-Shot Learning — Ekin Akyürek, Mehul Damani, Linlu Qiu, Han Guo, Yoon Kim, Jacob Andreas, 2024 https://scholar.google.com/scholar?q=The+Surprising+Effectiveness+of+Test-Time+Training+for+Few-Shot+Learning 5. Titans: Learning to Memorize at Test Time — Ali Behrouz, Peilin Zhong, Vahab Mirrokni, 2024 https://scholar.google.com/scholar?q=Titans%3A+Learning+to+Memorize+at+Test+Time 6. H2O: Heavy-Hitter Oracle: Efficient Generative Inference of Large Language Models with Heavy Hitters — Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Zhangyang Wang, Beidi Chen, et al., 2023 https://scholar.google.com/scholar?q=H2O%3A+Heavy-Hitter+Oracle%3A+Efficient+Generative+Inference+of+Large+Language+Models+with+Heavy+Hitters 7. SnapKV: LLM Knows What You are Looking for Before Generation — Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, Deming Chen, 2024 https://scholar.google.com/scholar?q=SnapKV%3A+LLM+Knows+What+You+are+Looking+for+Before+Generation 8. KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction — Jang-Hyun Kim, Jinuk Kim, Sangwoo Kwon, Jae W. Lee, Sangdoo Yun, Hyun Oh Song, 2025 https://scholar.google.com/scholar?q=KVzip%3A+Query-Agnostic+KV+Cache+Compression+with+Context+Reconstruction 9. Learning to Compress Prompts with Gist Tokens — Jesse Mu, Xiang Lisa Li, Noah Goodman, 2023 https://scholar.google.com/scholar?q=Learning+to+Compress+Prompts+with+Gist+Tokens 10. Cartridges: Lightweight and General-Purpose Long Context Representations via Self-Study — Sabri Eyuboglu, Ryan Ehrlich, Simran Arora, Neel Guha, Dylan Zinsley, Emily Liu, Will Tennien, Atri Rudra, James Zou, Azalia Mirhoseini, Christopher Ré, 2025 https://scholar.google.com/scholar?q=Cartridges%3A+Lightweight+and+General-Purpose+Long+Context+Representations+via+Self-Study 11. Fast KV Compaction via Attention Matching — Zweiger et al., 2026 https://scholar.google.com/scholar?q=Fast+KV+Compaction+via+Attention+Matching 12. The Power of Scale for Parameter-Efficient Prompt Tuning — Brian Lester, Rami Al-Rfou, Noah Constant, 2021 https://scholar.google.com/scholar?q=The+Power+of+Scale+for+Parameter-Efficient+Prompt+Tuning 13. Learning by Distilling Context — Charlie Snell, Dan Klein, Ruiqi Zhong, 2022 https://scholar.google.com/scholar?q=Learning+by+Distilling+Context 14. KVzap: Fast, adaptive, and faithful KV cache pruning — Jégou, Jeblick, 2026 https://scholar.google.com/scholar?q=KVzap%3A+Fast%2C+adaptive%2C+and+faithful+KV+cache+pruning 15. End-to-End Test-Time Training for Long Context — Tandon, Dalal, Li, Koceja, Rød, Buchanan, Wang, Leskovec, Koyejo, Hashimoto, Guestrin, McCaleb, Choi, Sun, 2025 https://scholar.google.com/scholar?q=End-to-End+Test-Time+Training+for+Long+Context 16. Prefix-Tuning: Optimizing Continuous Prompts for Generation — Li, Liang, 2021 https://scholar.google.com/scholar?q=Prefix-Tuning%3A+Optimizing+Continuous+Prompts+for+Generation 17. Adapting Language Models to Compress Contexts (AutoCompressors) — Chevalier, Wettig, Ajith, Chen, 2023 https://scholar.google.com/scholar?q=Adapting+Language+Models+to+Compress+Contexts+%28AutoCompressors%29 18. In-context Autoencoder for Context Compression in a Large Language Model — Ge, Hu, Wang, Wang, Chen, Wei, 2024 https://scholar.google.com/scholar?q=In-context+Autoencoder+for+Context+Compression+in+a+Large+Language+Model 19. Repeat After Me: Transformers are Better than State Space Models at Copying — Jelassi, Brandfonbrener, Kakade, Malach, 2024 https://scholar.google.com/scholar?q=Repeat+After+Me%3A+Transformers+are+Better+than+State+Space+Models+at+Copying 20. Efficient Streaming Language Models with Attention Sinks (StreamingLLM) — Xiao, Tian, Chen, Han, Lewis, 2024 https://scholar.google.com/scholar?q=Efficient+Streaming+Language+Models+with+Attention+Sinks+%28StreamingLLM%29 21. RULER: What's the Real Context Size of Your Long-Context Language Models? — Hsieh, Kriman, Acharya, Rekesh, Jia, Zhang, Ginsburg and others, 2024 https://scholar.google.com/scholar?q=RULER%3A+What%27s+the+Real+Context+Size+of+Your+Long-Context+Language+Models%3F 22. Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation — Feng, Lv, Cao, Xie, Zhou, 2025 https://scholar.google.com/scholar?q=Ada-KV%3A+Optimizing+KV+Cache+Eviction+by+Adaptive+Budget+Allocation 23. KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache — Liu, Yuan, Jin, Zhong, Xu, Braverman, Chen, Hu, 2024 https://scholar.google.com/scholar?q=KIVI%3A+A+Tuning-Free+Asymmetric+2bit+Quantization+for+KV+Cache Interactive Visualization: Latent Compaction with Test-Time Training: Learning Compact KV Caches

    Transcription

28. DepthBench: How Residual Connections Unlock Computational Depth in Transformers

    30 octobre 2026À suivre

    This episode examines DepthBench, a study of why stacking more transformer layers often adds little useful computation, and which residual and normalization designs fix that when parameter count and training …

    Notes de l’épisode

    This episode examines DepthBench, a study of why stacking more transformer layers often adds little useful computation, and which residual and normalization designs fix that when parameter count and training recipe are held constant. It reviews the Pre-LN "Curse of Depth" argument, in which growing residual-stream magnitude makes late layers act like identity. It then surveys ten architectures in three families: normalization tweaks such as LayerNorm Scaling and DeepNorm, multi-stream residuals such as Hyper-Connections and mHC, and cross-layer access such as Attention Residuals. The central claim is that the benefit of going deeper and narrower depends on the architecture: it can hurt under standard Pre-LN but keeps helping under two other designs. That would mean the common shallow-wide shapes are partly an artifact of the residual design. The discussion sets this against Kaplan-style scaling results, layer-pruning studies, and the rival explanation that deep models are simply mis-scaled in training. Listeners get a clear framework for "effective computational depth" and for judging whether the evidence supports the claim. Sources: 1. DepthBench: Measuring How Residual Connections Enable More Computational Depth — Keyu Wang, Yangyi Huang, Jiale Kang, David González-Martínez, Weiyang Liu, Shiwei Liu, 2026 http://arxiv.org/abs/2609.32534 2. Scaling Laws for Neural Language Models — Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, et al., 2020 https://scholar.google.com/scholar?q=Scaling+Laws+for+Neural+Language+Models 3. Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers — Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler, 2021 https://scholar.google.com/scholar?q=Scale+Efficiently%3A+Insights+from+Pre-training+and+Fine-tuning+Transformers 4. MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases — Zechun Liu, Changsheng Zhao, Forrest Iandola, et al. (Meta), 2024 https://scholar.google.com/scholar?q=MobileLLM%3A+Optimizing+Sub-billion+Parameter+Language+Models+for+On-Device+Use+Cases 5. The Curse of Depth in Large Language Models — Wenfang Sun, Xinyuan Song, Pengxiang Li, Lu Yin, Yefeng Zheng, Shiwei Liu, 2025 https://scholar.google.com/scholar?q=The+Curse+of+Depth+in+Large+Language+Models 6. Hyper-Connections (and follow-up: mHC, Manifold-Constrained Hyper-Connections) — Defa Zhu, Hongzhi Huang, Zihao Huang, et al. (ByteDance Seed) for HC; DeepSeek-AI researchers (Xie et al.) for mHC, 2024-2025 https://scholar.google.com/scholar?q=Hyper-Connections+%28and+follow-up%3A+mHC%2C+Manifold-Constrained+Hyper-Connections%29 7. Do Language Models Use Their Depth Efficiently? — Róbert Csordás, Christopher D. Manning, Christopher Potts, 2025 https://scholar.google.com/scholar?q=Do+Language+Models+Use+Their+Depth+Efficiently%3F 8. The Unreasonable Ineffectiveness of the Deeper Layers — Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso, Daniel A. Roberts, 2024 https://scholar.google.com/scholar?q=The+Unreasonable+Ineffectiveness+of+the+Deeper+Layers 9. ShortGPT: Layers in Large Language Models are More Redundant Than You Expect — Xin Men et al., 2024 https://scholar.google.com/scholar?q=ShortGPT%3A+Layers+in+Large+Language+Models+are+More+Redundant+Than+You+Expect 10. Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer (muP) — Greg Yang et al., 2022 https://scholar.google.com/scholar?q=Tensor+Programs+V%3A+Tuning+Large+Neural+Networks+via+Zero-Shot+Hyperparameter+Transfer+%28muP%29 11. Don't be lazy: CompleteP enables compute-efficient deep transformers — Nolan Dey et al., 2025 https://scholar.google.com/scholar?q=Don%27t+be+lazy%3A+CompleteP+enables+compute-efficient+deep+transformers 12. Scaling Laws with Model Shape / Gemstones: A Model Suite for Multi-Faceted Scaling Laws — Sean McLeish et al., 2025 https://scholar.google.com/scholar?q=Scaling+Laws+with+Model+Shape+%2F+Gemstones%3A+A+Model+Suite+for+Multi-Faceted+Scaling+Laws 13. The Impact of Depth on Compositional Generalization in Transformer Language Models — Jackson Petty et al., 2024 https://scholar.google.com/scholar?q=The+Impact+of+Depth+on+Compositional+Generalization+in+Transformer+Language+Models 14. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach — Jonas Geiping et al., 2025 https://scholar.google.com/scholar?q=Scaling+up+Test-Time+Compute+with+Latent+Reasoning%3A+A+Recurrent+Depth+Approach 15. Transformers, Parallel Computation, and Logarithmic Depth — Clayton Sanford, Daniel Hsu, Matus Telgarsky, 2024 https://scholar.google.com/scholar?q=Transformers%2C+Parallel+Computation%2C+and+Logarithmic+Depth 16. The Expressive Power of Transformers with Chain of Thought / The Parallelism Tradeoff — William Merrill, Ashish Sabharwal, 2023-2024 https://scholar.google.com/scholar?q=The+Expressive+Power+of+Transformers+with+Chain+of+Thought+%2F+The+Parallelism+Tradeoff 17. Resolving Discrepancies in Compute-Optimal Scaling of Language Models — Tomer Porian et al., 2024 https://scholar.google.com/scholar?q=Resolving+Discrepancies+in+Compute-Optimal+Scaling+of+Language+Models 18. DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging — Matteo Pagliardini et al., 2024 https://scholar.google.com/scholar?q=DenseFormer%3A+Enhancing+Information+Flow+in+Transformers+via+Depth+Weighted+Averaging 19. Peri-LN: Revisiting Layer Normalization in the Transformer Architecture — Jeonghoon Kim et al., 2025 https://scholar.google.com/scholar?q=Peri-LN%3A+Revisiting+Layer+Normalization+in+the+Transformer+Architecture Interactive Visualization: DepthBench: How Residual Connections Unlock Computational Depth in Transformers

    Transcription

29. Gradient Noise Scale Predicts Critical Batch Size for Large-Batch Training

    30 octobre 2026À suivre

    This episode examines the OpenAI paper on large-batch training, which asks why ImageNet training stalls at batches of tens of thousands of images while Dota 2 agents train on millions of observations. It …

    Notes de l’épisode

    This episode examines the OpenAI paper on large-batch training, which asks why ImageNet training stalls at batches of tens of thousands of images while Dota 2 agents train on millions of observations. It explains the tradeoff between serial optimizer steps and total examples processed. This tradeoff forms a Pareto frontier whose knee is the critical batch size, where compute efficiency is 50 percent. The central claim is that the gradient noise scale, a statistic computable from a single ordinary training run, predicts that critical batch size to within an order of magnitude. The discussion walks through the second-order loss derivation behind it, where progress per step is halved when batch size equals the noise scale. It also covers the theory's predictions: the noise scale grows during training, is larger for harder tasks and for reinforcement learning, and depends only weakly on model size. The authors admit the derivation rests on many unfounded assumptions. A listener would find it useful for understanding how to choose batch sizes, including why batch sizes that increase over training make sense, without brute-force sweeps. Sources: 1. An Empirical Model of Large-Batch Training — Sam McCandlish, Jared Kaplan, Dario Amodei, OpenAI Dota Team, 2018 http://arxiv.org/abs/1812.06162 2. An Empirical Model of Large-Batch Training — Sam McCandlish, Jared Kaplan, Dario Amodei, OpenAI Dota Team, 2018 https://scholar.google.com/scholar?q=An+Empirical+Model+of+Large-Batch+Training 3. Measuring the Effects of Data Parallelism on Neural Network Training — Christopher J. Shallue, Jaehoon Lee, Joseph Antognini, Jascha Sohl-Dickstein, Roy Frostig, George E. Dahl, 2018 https://scholar.google.com/scholar?q=Measuring+the+Effects+of+Data+Parallelism+on+Neural+Network+Training 4. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour — Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, Kaiming He, 2017 https://scholar.google.com/scholar?q=Accurate%2C+Large+Minibatch+SGD%3A+Training+ImageNet+in+1+Hour 5. Large Batch Training of Convolutional Networks (LARS) — Yang You, Igor Gitman, Boris Ginsburg, 2017 https://scholar.google.com/scholar?q=Large+Batch+Training+of+Convolutional+Networks+%28LARS%29 6. On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima — Nitish Shirish Keskar, Dheevatcha Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, Ping Tak Peter Tang, 2016 https://scholar.google.com/scholar?q=On+Large-Batch+Training+for+Deep+Learning%3A+Generalization+Gap+and+Sharp+Minima 7. Don't Decay the Learning Rate, Increase the Batch Size — Samuel L. Smith, Pieter-Jan Kindermans, Chris Ying, Quoc V. Le, 2017 https://scholar.google.com/scholar?q=Don%27t+Decay+the+Learning+Rate%2C+Increase+the+Batch+Size 8. Scaling Laws for Neural Language Models — Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, Dario Amodei, 2020 https://scholar.google.com/scholar?q=Scaling+Laws+for+Neural+Language+Models 9. Which Algorithmic Choices Matter at Which Batch Sizes? Insights From a Noisy Quadratic Model — Guodong Zhang, Lala Li, Zachary Nado, James Martens, Sushant Sachdeva, George E. Dahl, Christopher J. Shallue, Roger Grosse, 2019 https://scholar.google.com/scholar?q=Which+Algorithmic+Choices+Matter+at+Which+Batch+Sizes%3F+Insights+From+a+Noisy+Quadratic+Model 10. Three Factors Influencing Minima in SGD — Stanisław Jastrzębski, Zachary Kenton, Devansh Arpit, et al., 2017 https://scholar.google.com/scholar?q=Three+Factors+Influencing+Minima+in+SGD 11. Dota 2 with Large Scale Deep Reinforcement Learning — OpenAI, Christopher Berner, Greg Brockman, et al., 2019 https://scholar.google.com/scholar?q=Dota+2+with+Large+Scale+Deep+Reinforcement+Learning 12. Online Batch Selection / Adaptive sampling via gradient noise (Balles, Romero, Hennig) — Lukas Balles, Javier Romero, Philipp Hennig, 2016 https://scholar.google.com/scholar?q=Online+Batch+Selection+%2F+Adaptive+sampling+via+gradient+noise+%28Balles%2C+Romero%2C+Hennig%29 Interactive Visualization: Gradient Noise Scale Predicts Critical Batch Size for Large-Batch Training

    Transcription

30. SanSi: Looped Typed Decision Models for System 1.5 Thinking

    28 octobre 2026À suivre

    This episode examines SanSi, a small "typed decision" model that returns a probability for each caller-declared option (such as spam or not spam) in a single forward pass, with no generated text to parse. It …

    Notes de l’épisode

    This episode examines SanSi, a small "typed decision" model that returns a probability for each caller-declared option (such as spam or not spam) in a single forward pass, with no generated text to parse. It explains how the model gets more computation by looping the same 24-layer stack up to eight times, with a readout after every loop, so one model can serve any budget from one loop to eight. The discussion covers how this "System 1.5" approach differs from one-pass classifiers and from generated chain-of-thought. It also covers the earlier work it builds on: Universal Transformers, adaptive computation time, and pre-trained looped models such as Ouro. On training, the model uses a frozen backbone with LoRA adapters (about 61 million trainable parameters) and a loss averaged across all loops. Supervising only the final loop dropped loop-1 accuracy to 35.6%, against 58.4% with per-loop training. The episode also raises caveats: the Kahneman-style labels are framing rather than evidence, the RL case study uses the judge's own probability as reward rather than a truly verifiable one, and it is unclear whether looping beats spending the same compute another way. Sources: 1. SanSi: A Looped Typed Decision Model for System 1.5 Thinking — Shuyu Gan, Young-Jun Lee, Dongyeop Kang, 2026 http://arxiv.org/abs/2610.07730 2. SanSi: A Looped Typed Decision Model for System 1.5 Thinking — Shuyu Gan, Young-Jun Lee, Dongyeop Kang, 2026 https://scholar.google.com/scholar?q=SanSi%3A+A+Looped+Typed+Decision+Model+for+System+1.5+Thinking 3. Strictly Proper Scoring Rules, Prediction, and Estimation — Tilmann Gneiting, Adrian E. Raftery, 2007 https://scholar.google.com/scholar?q=Strictly+Proper+Scoring+Rules%2C+Prediction%2C+and+Estimation 4. On Calibration of Modern Neural Networks — Chuan Guo, Geoff Pleiss, Yu Sun, Kilian Q. Weinberger, 2017 https://scholar.google.com/scholar?q=On+Calibration+of+Modern+Neural+Networks 5. Generative Verifiers: Reward Modeling as Next-Token Prediction — Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, Rishabh Agarwal, 2024 https://scholar.google.com/scholar?q=Generative+Verifiers%3A+Reward+Modeling+as+Next-Token+Prediction 6. Universal Transformers — Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, Łukasz Kaiser, 2019 https://scholar.google.com/scholar?q=Universal+Transformers 7. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach — Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, Tom Goldstein, 2025 https://scholar.google.com/scholar?q=Scaling+up+Test-Time+Compute+with+Latent+Reasoning%3A+A+Recurrent+Depth+Approach 8. Scaling Latent Reasoning via Looped Language Models (Ouro) — Rui-Jie Zhu et al., 2025 https://scholar.google.com/scholar?q=Scaling+Latent+Reasoning+via+Looped+Language+Models+%28Ouro%29 9. Adaptive Computation Time for Recurrent Neural Networks — Alex Graves, 2016 https://scholar.google.com/scholar?q=Adaptive+Computation+Time+for+Recurrent+Neural+Networks 10. Multi-Scale Dense Networks for Resource Efficient Image Classification — Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger, 2018 https://scholar.google.com/scholar?q=Multi-Scale+Dense+Networks+for+Resource+Efficient+Image+Classification 11. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (Huginn) — Jonas Geiping et al., 2025 https://scholar.google.com/scholar?q=Scaling+up+Test-Time+Compute+with+Latent+Reasoning%3A+A+Recurrent+Depth+Approach+%28Huginn%29 12. PonderNet: Learning to Ponder — Andrea Banino, Jan Balaguer, Charles Blundell, 2021 https://scholar.google.com/scholar?q=PonderNet%3A+Learning+to+Ponder 13. Shallow-Deep Networks: Understanding and Mitigating Network Overthinking — Yigitcan Kaya, Sanghyun Hong, Tudor Dumitras, 2019 https://scholar.google.com/scholar?q=Shallow-Deep+Networks%3A+Understanding+and+Mitigating+Network+Overthinking 14. End-to-end Algorithm Synthesis with Recurrent Networks: Logical Extrapolation Without Overthinking — Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, Tom Goldstein, 2022 https://scholar.google.com/scholar?q=End-to-end+Algorithm+Synthesis+with+Recurrent+Networks%3A+Logical+Extrapolation+Without+Overthinking 15. Proper Scoring Rules / Strictly Proper Scoring Rules, Prediction, and Estimation — Tilmann Gneiting, Adrian Raftery, 2007 https://scholar.google.com/scholar?q=Proper+Scoring+Rules+%2F+Strictly+Proper+Scoring+Rules%2C+Prediction%2C+and+Estimation 16. Large Language Models Can Self-Improve / TTRL: Test-Time Reinforcement Learning — Yuxin Zuo et al. (TTRL); Jiaxin Huang et al., 2025 https://scholar.google.com/scholar?q=Large+Language+Models+Can+Self-Improve+%2F+TTRL%3A+Test-Time+Reinforcement+Learning 17. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena — Lianmin Zheng et al., 2023 https://scholar.google.com/scholar?q=Judging+LLM-as-a-Judge+with+MT-Bench+and+Chatbot+Arena 18. Chain-of-Thought Reasoning Without Prompting / Faithfulness work on CoT (Turpin et al., Language Models Don't Always Say What They Think) — Miles Turpin et al., 2023 https://scholar.google.com/scholar?q=Chain-of-Thought+Reasoning+Without+Prompting+%2F+Faithfulness+work+on+CoT+%28Turpin+et+al.%2C+Language+Models+Don%27t+Always+Say+What+They+Think%29 Interactive Visualization: SanSi: Looped Typed Decision Models for System 1.5 Thinking

    Transcription

Voici les 40 derniers épisodes sur 1 046. Les épisodes plus anciens se trouvent dans le flux du podcast.