[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-pivot-sd-diffusion-lm-self-distillation":3,"topics-all":38,"news-related-c448a366-372a-47dc-936f-8ea89103d238":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"c448a366-372a-47dc-936f-8ea89103d238","Pivot-SD:200道题让扩散LLM数学涨6分","扩散语言模型的后训练一直很粗:SFT 整段文本均权,RL 整条轨迹均摊。Pivot-SD 用信息增益找出去噪中真正决定答案走向的关键 token,只对它们奖惩:200 道题、4 次 rollout 就让 LLaDA-8B 数学与代码四项跑分全面超过 SFT 与在线 RL,耗时约为后者的八分之一。","扩散语言模型(dLM)想成为自回归之外的另一条路,但后训练一直是个粗活:SFT 把整段文本当教材,所有 token 一视同仁;在线 RL 把整条轨迹的奖励摊到每个位置。而 dLM 的去噪有个独特性质——少数几个早期承诺(commitment)一旦做出,剩余掩码位置的不确定性会骤降,整条回复的走向基本被这几个决定钉死。失败往往不是全盘皆错,而是毁在几个关键节点上。现有训练配方恰恰没有利用这个信号。\n\n## 只训「支点」token\n\n一个研究团队提出的 Pivot-SD 是离线自蒸馏框架,思路分两步:先用信息增益(information gain)度量每一次去噪承诺对剩余掩码位置不确定性的削减量,挑出 top-K 个高影响承诺作为「支点」(pivot);然后只对这些支点施加监督——正确轨迹的支点用交叉熵强化,错误轨迹的支点用 targeted unlikelihood 定点压制,失败轨迹里其余有效部分原样保留,不做无差别惩罚。这与 SFT 丢掉全部失败样本、RL 奖惩整条轨迹的粒度完全不同。\n\n## 200 道题的账本\n\n训练预算少得反常:200 道题、每题 4 次 rollout、1000 步。在 LLaDA-8B-Instruct 上(256 token 生成,三次运行取均值,与所有基线同预算),四项跑分全面超过 full-sequence SFT 与预算对齐的扩散 RL(diffu-GRPO、wd1++):\n\n- MATH500:基座 31.40 → 37.47,比 SFT-GT 的 35.47 高 2 个点\n- GSM8K:75.13 → 79.51\n- HumanEval+:32.32 → 40.43,四项中涨幅最大\n- MBPP+:43.12 → 47.12\n\n算力账更直接:800 条 rollout 一次性离线采样花 2.5 小时,训练本身 0.3 小时(2× NVIDIA L40),总墙钟约为在线 RL 的八分之一,平均准确率反超 3 分以上。论文已被 EMNLP 2026 收为主会 Oral([arXiv:2610.03665](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.03665))。\n\n## 值得注意的三点\n\n- **信号密度大于数据量。** 200 道题打赢整段 SFT,说明塑造模型行为的不是 token 数量,而是少数关键决策点。对验证器昂贵、标注昂贵的领域,这是条省钱的路线。\n- **失败样本终于被「外科手术式」利用。** 传统 SFT 直接丢弃错误轨迹,Pivot-SD 只惩罚导致错误的支点、保留其中有效的子步骤——负样本的价值第一次被精细定价。\n- **局限也明确。** 实验只覆盖 8B 量级、数学与代码任务、256 token 生成长度;信息增益选支点在超长生成和更大模型上是否依然高效,论文没有给出答案。\n\n当争论还聚焦在扩散 LM 能否追上自回归时,这篇论文提示了另一个方向:训练信号本身就该按信息量分配,而不是按 token 均摊。支点找到了,下一步是把它搬上更大的棋盘。\n\n参考:论文 [arXiv:2610.03665](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.03665) · 项目页 sunwoohong.github.io\u002Fpivot-sd","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.03665","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"0aed0b24-86a0-41fc-9f35-e3383c118031","en","Pivot-SD Trains Only the Pivots, Diffusion LM Gains +8","Pivot-SD trains only pivotal tokens in masked diffusion LMs: 200 questions lift LLaDA-8B past SFT and online RL at ~1\u002F8 the wall-clock cost.","Diffusion language models (dLMs) want to be the alternative to autoregressive generation, but their post-training has stayed crude: supervised fine-tuning treats every token of the final text as equally important teaching material, while online reinforcement learning spreads one trajectory-level reward across every position. Masked diffusion has a distinctive property here — a few early commitments during denoising sharply reduce the uncertainty over the remaining masked positions and effectively pin down the rest of the response. Failures are rarely wholesale; they trace back to a handful of critical decisions. Existing recipes ignore that signal entirely.\n\n## Train Only the Pivots\n\nPivot-SD is an offline self-distillation framework built around that signal. An information-gain metric measures how much each denoising commitment reduces uncertainty over the remaining masked block; the top-K highest-impact commitments become \"pivots\". Supervision is then applied to pivots only: pivots from successful trajectories are reinforced with cross-entropy, pivots from failed trajectories are suppressed with targeted unlikelihood, and the valid sub-steps inside failed trajectories are left untouched. That is a different granularity from SFT discarding failures wholesale or RL rewarding entire trajectories.\n\n## The 200-Question Ledger\n\nThe training budget is unusually small: 200 questions, four rollouts each, 1,000 steps. On LLaDA-8B-Instruct (256-token generation, mean of three runs, same budget as all baselines), Pivot-SD beats full-sequence SFT and budget-matched diffusion RL (diffu-GRPO, wd1++) on all four benchmarks:\n\n- MATH500: 31.40 base → 37.47, two points above SFT-GT's 35.47\n- GSM8K: 75.13 → 79.51\n- HumanEval+: 32.32 → 40.43, the largest gain of the four\n- MBPP+: 43.12 → 47.12\n\nThe compute ledger is just as direct: 800 rollouts sampled offline once for 2.5 hours plus 0.3 hours of training on 2× NVIDIA L40 — roughly one-eighth of the wall-clock of online RL, with average accuracy more than three points higher. The paper has been accepted as an EMNLP 2026 Main Conference Oral ([arXiv:2610.03665](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.03665)).\n\n## Three Takeaways\n\n- **Signal density beats data volume.** 200 questions outperforming full-sequence SFT says a model's behavior is shaped by a few decisive steps, not token count. For domains with expensive verifiers and annotations, this is a budget-friendly route.\n- **Failed trajectories finally get surgical use.** SFT throws them away; Pivot-SD penalizes only the pivots that caused the error and preserves the valid sub-steps around them — negative samples get a precise price for the first time.\n- **The limits are explicit.** Experiments cover an 8B model, math and code tasks, and 256-token generations. Whether pivot selection stays efficient at longer horizons and larger scales is left open.\n\nWhile the debate centers on whether diffusion LMs can catch up with autoregression, this paper points at a different lever: training signal should be allocated by information content, not spread evenly across tokens. The pivots are found — the next move is testing them on a bigger board.\n\nReference: paper [arXiv:2610.03665](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.03665) · project page sunwoohong.github.io\u002Fpivot-sd","pivot-sd-diffusion-lm-self-distillation","2026-10-05T17:09:21Z","2026-10-05T17:09:26.517728Z","2026-10-05T17:09:26.517738Z",true,"agent",641,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"a78b61ba-6899-4b4f-86b0-8d6c1753308b","亚马逊 ALoDLM:扩散语言模型学会按难度分配算力","alodlm-adaptive-looped-diffusion","2026-10-07T13:15:53+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"8c579d8d-4114-4d7d-92c1-22f5fab0537b","连续流语言模型跑到 8B：Lafa 单模型自验证解码提速 2.54 倍","large-flow-language-models-lafa-self-speculative-decoding","2026-10-21T03:09:03+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"b369ce1a-f8a3-4aba-ba78-c73c097a413d","MemT-1.0：记忆变成模型内置能力，128K 历史存储降 66 倍","memt-1-0-memory-native-foundation-model","2026-10-20T11:07:23+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"6ede1566-74c5-4bdd-9dd5-6d3f118db876","投机解码的重复陷阱:D-Loop 两轮循环破局","d-loop-diffusion-speculative-decoding","2026-10-06T15:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"34d309c3-5757-469a-85ff-b524c8865037","Queen 模型边下棋边解说:七轮自蒸馏涨 900 Elo","queen-chess-language-model-self-distillation","2026-10-05T23:04:34+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"c361090b-291c-4357-9bd3-54845025f29d","CaRE-KD 让蒸馏学会看「教师脸色」:NeurIPS 2026 选中 IIT Delhi","care-kd-confidence-distillation-neurips-2609-36734","2026-10-05T03:00:00+00:00"]