[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-partrep-prompt-repetition-nll":3,"topics-all":33,"news-related-88e8d882-dbf3-4912-b4b3-68f795e9e82e":52},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":20,"news_slug":26,"published_at":27,"created_at":28,"modified_at":29,"is_published":30,"publish_type":31,"image_url":13,"view_count":32},"88e8d882-dbf3-4912-b4b3-68f795e9e82e","PARTREP：用 NLL 选难 token，KV 占用压到 59%","Decoder-only LLM 的因果注意力天然造成一种信息流不对称——靠前的 token grounding 弱,靠后的 token grounding 强。最朴素的补救办法是\"把 prompt 整体复制一份贴到末尾\"(prompt repetition),让原本靠前的 token 获得一次\"晚到\"的机会。问题是代价太重:它让 KV cache 翻倍,prefill 注意力变成 4 倍,在长上下文场景里基本不可用。\n\n7 月 2 日挂在 arXiv 的 PARTREP (2607.01792) 走了一条更精打细算的路:不是\"复不复读\"二选一,而是只复读\"难 token\"。判定\"难\"的标准是 token-wise 负对数似然(NLL)——模型自己预测越不准的 token,在前段越难从上下文里被还原,放在末尾复读一次的边际收益就越大。\n\n真正让它工程化的是打分本身。为了不在选 token 之前再跑一遍完整 forward,作者训练了一个轻量级 gate,用前几层的 hidden state 就能预测高 NLL token,选择动作发生在 prefill 中段的 early exit。论文在 8 个 benchmark(MMLU、GSM8K、RULER 等)上、跨 3 个模型族(Qwen2.5、Llama3.2、Gemma4)做了验证:用 59.4% 的 KV cache + 79% 的 prefill FLOPs,就能拿到接近 full repetition 的增益。\n\nPARTREP 不是新架构,也不是新训练范式,而是把\"在已经能跑的系统上插一个聪明判断\"做到了工业可用的程度。它和 self-consistency、speculative decoding 同属\"借模型自身不确定性做决策\"一脉——差别在于 PARTREP 完全不动训练,落地路径最短。真正的取舍在于:这个 learned gate 要为每个目标模型重新训,跨模型的可移植性是下一篇工程论文该回答的问题。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.01792","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[21],{"id":22,"lang":23,"title":24,"summary":25,"content":13},"6506ec23-3748-41a7-9d35-b4e5fe79b649","en","PARTREP: NLL-selected hard tokens, KV usage down to 59%","The causal attention of Decoder-only LLMs naturally creates an asymmetric information flow — tokens earlier have weak grounding, tokens later have strong grounding. The most straightforward remedy is \"copy the prompt and paste it at the end\" (prompt repetition), giving originally early tokens a \"late\" chance. The problem is the cost is too heavy: it doubles the KV cache, and prefill attention becomes 4×, basically unusable in long-context scenarios. PARTREP (2607.01792), posted to arXiv on July 2, takes a more meticulous path: not \"whether to repeat\" as a binary choice, but only repeat \"hard tokens\". The criterion for \"hard\" is token-wise negative log-likelihood (NLL) — the more uncertain the model's own prediction for a token, the harder it is to be restored from context in the front segment, and the higher the marginal benefit of repeating once at the end. What really industrializes it is the scoring itself. To avoid running another full forward before selecting tokens, the authors train a lightweight gate, using the hidden state of the first few layers to predict high-NLL tokens, with the selection action happening at the early exit in the middle of prefill. The paper validates on 8 benchmarks (MMLU, GSM8K, RULER, etc.), across 3 model families (Qwen2.5, Llama3.2, Gemma4): with 59.4% KV cache + 79% prefill FLOPs, you can get close to the gain of full repetition. PARTREP is neither a new architecture nor a new training paradigm, but a smart judgment inserted into an already-run system, taken to an industry-usable degree. It's in the same vein as self-consistency and speculative decoding, \"making decisions with the model's own uncertainty\" — the difference is PARTREP doesn't touch training at all, the landing path is the shortest. The real trade-off is: this learned gate needs to be retrained for each target model, cross-model portability is a question for the next engineering paper.","partrep-prompt-repetition-nll","2026-07-03T10:09:00Z","2026-07-03T10:10:32.346818Z","2026-08-19T02:08:40.142862Z",true,"agent",188,[34,43],{"slug":35,"tag_slug":35,"title_zh":36,"title_en":37,"intro_zh":38,"intro_en":39,"id":40,"is_active":30,"created_at":41,"modified_at":42},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":44,"tag_slug":44,"title_zh":45,"title_en":46,"intro_zh":47,"intro_en":48,"id":49,"is_active":30,"created_at":50,"modified_at":51},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":53},[54,59,64,69,74,79],{"id":55,"title":56,"news_slug":57,"published_at":58},"95e663a4-2cdf-454c-9787-b154fbd41909","TokenRouter:token 级路由提速 64 倍","tokenrouter-token-level-llm-routing","2026-10-09T21:11:28+00:00",{"id":60,"title":61,"news_slug":62,"published_at":63},"44b118b2-cb3d-47bb-8f26-bbce777cfb31","1 亿 rollout 背后:Beam 的 RL 训练工厂","beam-rl-factory-100m-rollouts","2026-10-07T07:15:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"ccce6dfe-776f-4cd8-9605-9163daea4627","Jeff 开源决策模型:2B 追平 Jev","jeff-open-decision-models-0-8b","2026-09-29T15:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"523d3ae4-9c60-4bcb-8e6c-ea98e58c7f71","一次前向一个决策:GLM-5.3-Flash 平替 Jev","glm-flash-single-token-decisions","2026-09-28T13:11:55+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"bc642ebf-9b1c-41cf-98d1-dd55393fd429","HeadWiseKV:无训练KV cache压缩让混合LLM长上下文从114K推到161K","headwisekv-training-free-kv-cache-compression","2026-09-03T03:44:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"f65e204c-0115-4b50-9113-2c3bb2ff6637","ReCache:给 Agent 的工具记忆装上独立缓存,KV 内存砍 92%、首 token 提速 3.655 倍","recache-agent-kv-cache-reuse","2026-08-24T15:30:00+00:00"]