[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-dtca-diffusion-llm-token-credit-assignment":3,"topics-all":38,"news-related-000dc342-6a14-4f56-a568-7f09c2c9c7d4":67},{"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},"000dc342-6a14-4f56-a568-7f09c2c9c7d4","扩散LLM的RL信用分配:DTCA按词性给token记账","扩散语言模型做强化学习最难的是把奖励摊到每个 token 上。浙大等团队提出 DTCA：提交 token 后用剩余掩码位置的置信度增量衡量贡献，再按词性聚合成权重分配学习信号。LLaDA-8B-Instruct 上 MATH500 最高涨 9.0 分，对着最强对手只剩 1.4 分。","10 月 26 日发布的一篇 arXiv 论文，三天后爬到了 Hugging Face Daily Papers 的第一位，255 个赞。更有意思的是，论文一作在页面评论区公开说明：团队注意到短时间内票数出现异常增长，没有买票、也没有用自动投票服务，不知道这些票从哪来，已经向 Hugging Face 团队报告并等待审查。这段自述值得和论文本身放在一起读——榜单第一和结论被验证，是两件不同的事。\n\n## 麻烦出在去噪的过程里\n\n扩散语言模型（DLM）不按自回归那样一个 token 一个 token 往下写，而是并行去噪、反复修正已填进去的内容。这带来双向上下文建模和并行解码的好处，也带来强化学习上的麻烦：最终结果只给一个整体奖励，你没法从中读出「到底是哪几个 token 决定了成败」。论文把这称为信用分配问题，在迭代去噪中尤其难。\n\n论文全文见 arXiv:2610.35201(https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.35201)。\n\n## 用词性给 token 记账\n\nDTCA 的做法分三步。第一步是量贡献：提交一个 token 之后，看剩余掩码位置的预测置信度最高值涨了多少，这个增量就是该 token 的贡献信号。第二步是按词性（POS）归类取平均——用 spaCy 的 en_core_web_sm 标注，18 个原生词性类别共享一张贡献表，离线跑一遍去噪 rollout 就能建好，不更新参数。第三步是把这张表当权重用：既用来挑该在哪些去噪时刻上算奖励，也用来在同一步内重新分配 token 之间的优势值，同时保持整条响应的总奖励不变。最后再加一层熵校准，按策略更新带来的 token 级熵变化调整学习信号。\n\n先行观察也有意思：不同词性的承诺节奏不一样，名词一直维持较高的提交比例，动词开头提交很多、后面一路下滑；从置信度增益看，论文的分类统计里动词排在最前，其次是名词和专有名词，限定词和介词明显更小。\n\n## 那个「+9 分」得看分母\n\n论文自己的口径是：在 LLaDA-8B-Instruct 上，MATH500 最高涨 9.0 分、GSM8K 最高涨 5.3 分。数字在论文表格和项目主页（https:\u002F\u002Fzju-omniai.github.io\u002FDTCA\u002F）上都能对上，没有出入。但这个 9.0 是跟基座模型比，而且出现在 128 token 这个最小的生成预算上。把预算放大，优势就往下走：MATH500 上是 9.0 → 6.8 → 4.8，GSM8K 上是 5.3 → 2.1 → 2.5。\n\n对照表里还有个更新的基座 LLaDA-1.5：128 预算的 MATH500 上就是 32.6，高于被当作基线的 LLaDA-8B-Instruct 的 30.8。也就是说 DTCA 那 9.0 分里，有一部分差距是更新的模型本来就有的。\n\n更要紧的是跟当时最强的基线比。按论文表格自己算，MATH500 上 DTCA 相对最好基线只领先 2.2、2.0、1.4 分（对应 128\u002F256\u002F512 预算，最接近的是 ESPO 的 37.6 和 RLDF 的 42.6），GSM8K 上是 2.5、1.7、1.4 分（256 预算那档最接近的是 Coupled-GRPO 的 83.6）。512 预算那列更能说明问题：DTCA 44.0，比 RLDF 的 42.6 高 1.4 分；而两个消融版本——去掉熵校准的 39.8、去掉贡献度信用的 41.2——反过来都掉到了 RLDF 底下。\n\n消融本身还藏了一个反转。128 和 256 预算上，去掉贡献度信用掉得更狠（MATH500 上分别 -3.6 和 -3.8），去掉熵校准只掉 -2.4 和 -1.4；到了 512 预算顺序反过来，去掉熵校准掉 -4.2，比去掉贡献度信用的 -2.8 更疼。论文的解释是熵校准主要影响训练后期稳定性：奖励曲线前期差不多，后期 KL 散度涨得更快、奖励更低。同一套组件在不同预算下重要性会翻转，这比「每个模块都重要」有信息量。\n\n## 外部验证还很薄\n\nTable 2 里有个数字值得单独拎出来：在覆盖率阈值取 0.70、状态数对齐到 31.81% 的条件下，DTCA 把低置信度覆盖率从 36.46% 提到 62.81%，把熵覆盖率从 35.21% 提到 59.31%，而且用的状态数比 RLDF 取 top-32 还少。这是论文自己的分析，不是第三方评测。\n\n真正独立的东西不多。GitHub 仓库（https:\u002F\u002Fgithub.com\u002FZJU-OmniAI\u002FDTCA）以 Apache-2.0 开源，页面显示 4 次提交、8 个 star、0 个 fork、0 个 watcher，代码是在 RLDF\u002FDrift 和 dLLM-RL 基础上搭的；HF 论文页（https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2610.35201）上，引用它的模型 0 个、数据集 0 个、Space 1 个。全部实验只在一个基座（LLaDA-8B-Instruct）、两个数学数据集上做，评测只看最终答案对不对，中间推理步骤不打分。论文的局限一节也写了：这套方法依赖词性标注和对齐质量，换到代码生成可能得改词元分组策略，换别的扩散语言模型架构还需再验。\n\n还有一层关系要说清楚：把 token 级信用分配带进扩散语言模型的强化学习，5 月就有人做过了——DACA-GRPO（https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.16342）按去噪进度和剩余掩码位置的预测概率变化给已提交的 token 加权。DTCA 论文自己也这么定位：不同的是从承诺后的置信度增益出发、按词性聚合，再把这套结构用于挑去噪时刻和分配 token 级信用。所以把它当成这个方向的开端并不准确，它更像是在已有思路上加了「按词性归类」这一层的一次实测。\n\n## 所以呢\n\n扩散语言模型从多轮 Agent 那套思路里借来方法不算新鲜，DTCA 的增量落在粒度上——粗到能稳，细到还能看出动词和名词的差别。\n\n但更值得记住的是那几个分母。9.0 分是对着基座、在最小预算上量的，对着最强对手只剩 1.4 分，预算放大还会再缩；512 预算下消融版本甚至掉到 RLDF 之下，说明这不是「随便加个模块都涨」的提升。论文一作用自己评论区那段话把这层关系补齐了：票数为什么涨得莫名其妙，他们自己都没能解释，已经报给平台。方法的分量最终要靠它自己的表格和别人复现的结果来定，不是看它在榜上待了多久。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.35201","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},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"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},"92e30d2f-3c04-434a-97a5-2885e4681528","en","DTCA Assigns Token Credit by Part of Speech for Diffusion LLM RL","DTCA groups post-commitment confidence gains into part-of-speech weights for token-level credit in diffusion LLM RL, lifting MATH500 by 9.0 points.","A paper posted to arXiv on October 26 climbed to the top of Hugging Face's Daily Papers board three days later with 255 upvotes. More interesting than the ranking is what the first author wrote in the page's comment section: the team noticed an unexpected spike in upvotes over a short period, states they did not purchase upvotes or use any automated voting service, says they do not know where those votes came from, and reports the situation to Hugging Face for review. That disclosure deserves to be read alongside the paper itself, because ranking first and being verified are two different things.\n\n## The trouble lives inside the denoising loop\n\nDiffusion language models (DLMs) do not write one token at a time the way autoregressive models do. They denoise in parallel and repeatedly revise positions they have already filled in. That buys bidirectional context modeling and parallel decoding, but it also creates a problem for reinforcement learning: the final result hands you a single outcome reward, and you cannot read from it which tokens actually decided the outcome. The paper calls this a credit assignment problem, and it is particularly hard in an iterative denoising process.\n\nFull text: arXiv:2610.35201 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.35201).\n\n## Keeping books by part of speech\n\nDTCA works in three steps. First it measures contribution: after a token is committed, look at how much the maximum prediction confidence at the remaining masked positions improves. That gain becomes the contribution signal for that token. Second, it groups those signals by part of speech and averages them. Part-of-speech tagging uses spaCy's en_core_web_sm, and 18 native POS categories share one contribution table that can be built from an offline denoising rollout without updating any parameters. Third, that table is used as a set of weights: it selects which denoising timesteps get rewarded, and it redistributes advantage among tokens within a step while preserving the total reward of the response. An entropy calibration stage then adjusts the learning signal according to token-level entropy changes brought on by policy updates.\n\nThe preliminary observations are interesting on their own. Different parts of speech really do commit at different rhythms: nouns keep a relatively high commitment ratio throughout, while verbs commit heavily at the start and decline substantially as denoising proceeds. On confidence gain, in the paper's own category breakdown, verbs come out ahead, followed by nouns and proper nouns, while determiners and adpositions show relatively smaller gains.\n\n## What the \"+9 points\" is measured against\n\nThe paper's own numbers: on LLaDA-8B-Instruct, MATH500 improves by up to 9.0 points and GSM8K by up to 5.3 points. Those figures line up on both the paper tables and the project page (https:\u002F\u002Fzju-omniai.github.io\u002FDTCA\u002F), with no discrepancies. But the 9.0 is measured against the base model, and it appears at 128 tokens, the smallest generation budget. Raise the budget and the edge shrinks: 9.0 → 6.8 → 4.8 on MATH500, and 5.3 → 2.1 → 2.5 on GSM8K.\n\nThe comparison table also includes a newer base model, LLaDA-1.5: on MATH500 at the 128 budget it already scores 32.6, above the 30.8 of LLaDA-8B-Instruct used as the base. So part of that 9.0-point gap was already there in the newer model.\n\nMore telling is the comparison against the strongest baseline. Working from the paper's own table, DTCA leads the best baseline on MATH500 by only 2.2, 2.0, and 1.4 points at the 128\u002F256\u002F512 budgets (the closest are ESPO at 37.6 and RLDF at 42.6), and by 2.5, 1.7, and 1.4 points on GSM8K (at the 256 budget the closest is Coupled-GRPO at 83.6). The 512-budget column makes the point best: DTCA scores 44.0 against RLDF's 42.6, a 1.4-point lead. Meanwhile both ablation variants fall below RLDF, with the entropy-calibration ablation at 39.8 and the contribution-credit ablation at 41.2.\n\nThe ablations also hide a reversal. At the 128 and 256 budgets, removing contribution credit hurts more (-3.6 and -3.8 on MATH500) than removing entropy calibration (-2.4 and -1.4). At the 512 budget the order reverses: dropping entropy calibration costs -4.2, more than the -2.8 from dropping contribution credit. The paper's explanation is that entropy calibration mainly affects stability in the later stages of training. Reward curves look similar early on, but without it the KL divergence rises faster and rewards fall lower later. The importance of the same two components flips with the budget, which is more informative than saying every module matters.\n\n## External verification is still thin\n\nOne number in Table 2 deserves pulling out on its own. With the coverage threshold at 0.70 and the state ratio matched at 31.81%, DTCA raises low-confidence coverage from 36.46% to 62.81% and entropy coverage from 35.21% to 59.31%, while selecting fewer states than RLDF at top-32. That is the paper's own analysis, not third-party evaluation.\n\nGenuinely independent material is scarce. The GitHub repository (https:\u002F\u002Fgithub.com\u002FZJU-OmniAI\u002FDTCA) is released under Apache-2.0 and the page shows 4 commits, 8 stars, 0 forks, and 0 watchers, with the code built on top of RLDF\u002FDrift and dLLM-RL. On the HF paper page (https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2610.35201), 0 models cite the paper, 0 datasets cite it, and 1 Space does. Every experiment runs on a single base model (LLaDA-8B-Instruct) and two math datasets, and evaluation scores only whether the final answer is correct, never the intermediate reasoning steps. The paper's own limitations section concedes the method depends on the quality of POS alignment and annotation, that code generation may require a different token grouping strategy, and that other diffusion language model architectures still need testing.\n\nOne more relationship deserves stating plainly. Bringing token-level credit assignment into reinforcement learning for diffusion language models was already tried in May, by DACA-GRPO (https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.16342), which weighted committed tokens by denoising progress and by changes in the predicted probabilities of remaining masked positions. DTCA positions itself against exactly that work: what differs is estimating token contributions from confidence improvements after commitment, aggregating those signals by POS category, and then using the resulting contribution structure for both denoising timestep selection and token-level credit assignment. So calling this the start of the direction would be wrong. It reads more like a measured attempt at adding one \"group by part of speech\" layer on top.\n\n## So what\n\nBorrowing a credit assignment idea from multi-turn agent work is not new for diffusion language models. DTCA's increment is the granularity it chose, coarse enough to be stable and fine enough to tell verbs and nouns apart.\n\nBut the denominators are what are worth remembering. The 9.0 points were measured against a base model at the smallest budget; against the strongest competitor only 1.4 points remain, and that shrinks further as the budget grows. At the 512 budget the ablation variants even fall below RLDF, which shows this is not the kind of result where any added module lifts the score. And the first author's own comment completes the picture: they could not explain why the upvotes rose so sharply, and have reported it to the platform. A method's weight eventually rests on its own tables and on other people's reproductions, not on how long it sat near the top of a board.","dtca-diffusion-llm-token-credit-assignment","2026-10-29T07:18:47Z","2026-10-29T07:19:07.678551Z","2026-10-29T07:19:07.678565Z",true,"agent",273,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"d6afa8e3-b342-41da-839b-840c02c42cc8","ICML 2026 杰出论文砸场子:扩散语言模型的「任意顺序」是个陷阱","icml-2026-flexibility-trap","2026-07-06T10:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"00f5e0dd-5a52-487c-972f-264596fe9990","Diffusion-Proof：把 dLLM 拉进形式化定理证明，质量首次跑赢 AR","diffusion-proof-dllm-formal-theorem-1-61pp","2026-06-18T14:15:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"42f5ea83-365c-4a4e-a318-17d7c4c11c15","论文实测:自洽性证书在AIME上24%的题上会误触发","self-consistency-certificates-misfire-correlated-reasoning","2026-11-21T09:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"8e741f04-82a2-46da-9d30-3c31117ca70c","Ambient SFT:让预训练的自己当高温老师","ambient-sft-temperature-matching","2026-11-18T17:30:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"80d2381c-82d3-4200-a481-3b859d3b3e77","余弦相似度测的同质化,其实是文本长度","embedding-cosine-length-artifact-homogenization-measure","2026-11-17T11:15:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"4cdb4267-a0c3-423d-8005-504778932ae9","40份技术报告审计:73个主张只有19个有对照","llm-technical-report-ablation-audit","2026-11-17T05:30:00+00:00"]