[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-t2quant-task-aware-diffusion-llm-quantization":3,"topics-all":38,"news-related-4b73ca86-414e-454b-b2bf-df0d60e18d65":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},"4b73ca86-414e-454b-b2bf-df0d60e18d65","扩散LLM量化:T2Quant按去噪时刻分配精度","扩散LLM量化不能一刀切：T2Quant 让精度按去噪时刻分配，同预算最高涨 6.7 分。Technion 团队发现量化敏感度既随任务变也随去噪步变，GSM8K 破坏集中在最初 8 步，HumanEval 却在中段。该方法用无标签信号加动态规划，在三个扩散 LLM、六个基准上把平均激活精度压在 4.00 bit。","扩散语言模型（DLM）不是逐词解码，而是反复擦除重写，每一步去噪只提交一部分 token。这让量化很棘手——自回归那套「统一压到 4-bit」搬过来，常在关键的去噪窗口上翻车。\n\nTechnion 与本-古里安大学 INSIGHT Lab 团队在 arXiv 上给出 T2Quant，把量化精度从「一个全局数字」改成「沿去噪时间轴分配的一份预算」（[论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.31198)）。\n\n## 敏感度既看任务，也看时刻\n\n三个被测模型的曲线形状截然不同。GSM8K 极度前置，峰值落在最初 8 步；MMLU-Pro 偏早但更宽，前四个窗口（第 0–31 步）都在峰值 10% 以内；HumanEval 相反，前期几乎不敏感，峰值宽泛落在第 40–63 步。项目页（[T2Quant](https:\u002F\u002Fofekavivi.github.io\u002FT2Quant\u002F)）读数更直观：沿 W4A4 轨迹，GSM8K 第一个窗口就承载 27% 的任务相关偏移，是均匀分配的 4.4 倍。这套形状也稳定：LLaDA-1.5 与 LLaDA-8B-Instruct 的逐窗口相关系数为 0.99、0.98、0.92。\n\n## 不需要标签的「该加精度」信号\n\nT2Quant 不去新造量化器：不改权重、不改采样器、不改底层 quantizer，只决定每个去噪窗口跑几比特，宽度取自 {3, 4, 6, 8}，平均压在 4.00 bit。\n\n难点在于没有标准答案、也没有任务专属指标时，怎么判断哪个窗口值得加精度。解法分两步：先用全精度模型生成参考轨迹，把保存的状态在候选宽度下重放，得到该窗口的量化散度；再乘上「下游影响」——这一步提交的 token 把后续未解码位置的信念推动了多少。两项相乘得到边际收益，再用动态规划选出最优宽度。\n\n「下游影响」不能省：只看散度会误判 HumanEval 的最后一个窗口，那几步提交的 token 影响不了后面任何东西。\n\n## 数字\n\n三个扩散 LLM（Dream-7B-Instruct、LLaDA-1.5-8B、LLaDA-8B-Instruct）、六个基准（GSM8K \u002F HumanEval \u002F ARC-Challenge \u002F MMLU-Pro \u002F DROP \u002F BBH）上，4-bit 权重、平均激活 4.00 bit：\n\n- 9 个「基线 × 模型」组合的平均精度全部上升，54 个基准级对比里 42 个改善\n- 最大增益在 Dream 上达 6.7 个百分点，LLaDA-1.5 为 3.4，LLaDA-8B-Instruct 为 3.5；最好的量化配置相对最强独立基线再领先 3.55 \u002F 1.13 \u002F 1.40 个点\n- Dream 上 GSM8K 从 35.6 涨到 53.5（DLLMQuant）、从 46.1 涨到 58.8（STaR-Quant）\n- 同为 4.00 bit 预算的随机调度，在 Dream 上比原基线低 14–19 个点；RTN 在 Dream 上直接崩到 0.00\n\n消融指向同一结论：16 条无标签 prompt 就能拿到约 75% 的收益；LLaDA-8B-Instruct 的 GSM8K 上，统一 W4A4 是 44.1、T2Quant 是 49.4、FP16 是 52.9。\n\n## 几点保留\n\n论文自己也标了边界：硬件效率完全没测，量化是 bf16 算术下的模拟；精度调度与任务相关，换任务要重新标定；量化本身可能改变安全相关的行为。而 T2Quant 只是能插在任意后训练量化方法之上的调度层。\n\n## 结论\n\n自回归时代的习惯是「精度是一个全局常数」。扩散 LLM 打破了这个前提——不同去噪时刻的重要性可以差出数倍。把精度改成按时刻分配的预算，更值得推广。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.31198","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"419ada23-6082-46bc-8ced-2f0304d19a04","en","T2Quant: Task-Aware Quantization for Diffusion LLMs","T2Quant schedules quantization precision across the denoising timeline of diffusion LLMs, gaining up to 6.7 points on three models and six benchmarks.","Diffusion language models (DLMs) do not decode left to right. They start from a fully masked block and iteratively unmask tokens over many denoising steps, which creates a quantization problem that simply does not exist in autoregressive models. A team from Technion and the INSIGHT Lab at Ben-Gurion University argues that the default inherited approach of uniform bit allocation wastes precision in precisely the wrong places. Their framework, T2Quant, treats quantization precision as a resource to be scheduled over the denoising timeline rather than a single global constant (paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.31198).\n\n## Sensitivity depends on the task and on the moment\n\nThe paper's first finding is that activation quantization sensitivity in DLMs is not uniform. It varies along two axes at once, and the two axes turn out to matter independently.\n\nAlong the task axis, the three evaluated models produce three very different shapes. GSM8K is strongly front-loaded, with its peak inside the first 8 denoising steps. MMLU-Pro is also early but broad: its first four windows, covering steps 0 to 31, all sit within 10 percent of its peak. HumanEval behaves in the opposite way, barely sensitive early and peaking broadly across the middle of the trajectory, roughly steps 40 to 63. The project page puts the same finding in more concrete terms: along the W4A4 trajectory of LLaDA-8B-Instruct, GSM8K's very first window carries 27 percent of the task-relevant divergence, 4.4 times what uniform spreading would give, and the first quarter of the trajectory absorbs 55 percent. HumanEval is different again, with five mid-trajectory windows spanning steps 32 to 71 together accounting for 52 percent.\n\nThese profiles also turn out to be stable. Comparing the per-window curves of LLaDA-1.5 against LLaDA-8B-Instruct gives Pearson correlations of 0.99 on GSM8K, 0.98 on HumanEval and 0.92 on MMLU-Pro, meaning the temporal profile survives preference alignment.\n\n## A label-free signal for where precision pays\n\nMethodologically, T2Quant deliberately avoids inventing a new quantizer. It does not touch the weights, the sampler, or the underlying quantizer. It only decides which bit width each denoising window runs at, drawing from the set {3, 4, 6, 8} bits while holding the average strictly at 4.00 bits.\n\nThe hard part is deciding which window deserves more precision without access to ground-truth labels or task-specific metrics. The paper solves this in two steps. First it generates a full-precision reference trajectory per calibration prompt and replays every saved state under each candidate width, isolating the local quantization effect and yielding a divergence measure for that window. Second it weights that divergence by downstream influence, defined as how much the tokens committed in that window actually move the model's beliefs about positions that are still masked. Multiplying the two gives an estimate of the marginal benefit of upgrading that window, and a dynamic program over windows and cumulative cost then selects the globally optimal allocation under the budget.\n\nThe influence term is not optional. Divergence alone misfires on HumanEval's final window, whose committed tokens cannot affect anything that follows.\n\n## The numbers\n\nEvaluated on three masked-diffusion models, Dream-7B-Instruct, LLaDA-1.5-8B and LLaDA-8B-Instruct, across six benchmarks including GSM8K, HumanEval, ARC-Challenge, MMLU-Pro, DROP and BBH, with 4-bit weights and average activation precision fixed at 4.00 bits:\n\n- Average accuracy improves in all nine baseline-model combinations, with 42 of 54 benchmark-level comparisons improving. GSM8K and ARC-Challenge improve in every one of the nine combinations.\n- Gains over the corresponding baselines reach 6.7 percentage points on Dream, 3.4 on LLaDA-1.5 and 3.5 on LLaDA-8B-Instruct.\n- The best quantized configuration leads the strongest standalone baseline by a further 3.55, 1.13 and 1.40 points, leaving gaps to FP16 of only 2.0, 2.5 and 1.6 points.\n- On Dream, GSM8K accuracy climbs from 35.6 to 53.5 with DLLMQuant and from 46.1 to 58.8 with STaR-Quant. The other two models see GSM8K gains of 9.6 and 10.4 points.\n- The controls are the striking part. Random schedules under the same 4.00-bit average budget land 14 to 19 points below the original baselines on Dream, while RTN, which skips both rotation and GPTQ, collapses to 0.00.\n\nThe ablations point the same way. Sixteen unlabeled calibration prompts already recover about 75 percent of the gain. On LLaDA-8B-Instruct and GSM8K, uniform W4A4 scores 44.1, T2Quant scores 49.4 and FP16 scores 52.9. Squeeze the budget down to an average of 3.75 bits and T2Quant still reaches 48.4, beating uniform W4A4 at a full 4 bits. Pushing precision beyond 4 bits, by contrast, adds essentially nothing.\n\n## Where the caveats are\n\nThe paper flags its own limits. Hardware efficiency, meaning inference speed and energy consumption, is not measured at all, and all quantization is simulated in bf16 arithmetic. The precision schedule is task-dependent, so a new task means recalibration, and calibration needs representative unlabeled prompts from that task. Quantization itself can also alter safety-related behavior, which the authors list as future work.\n\nFor DLMs specifically, all three are separable. T2Quant is a scheduling layer that can sit on top of any post-training quantization method, binding to no particular quantizer and leaving the sampler alone.\n\n## The takeaway\n\nThe habit inherited from the autoregressive era is that precision is a global constant. DLMs break that premise. With the same weights and the same generation run, the importance of different denoising moments can differ by a factor of several. Turning precision from one number into a budget allocated over time is arguably the more transferable idea here, rather than just one more quantization trick.","t2quant-task-aware-diffusion-llm-quantization","2026-10-24T01:24:29Z","2026-10-24T01:30:27.726185Z","2026-10-24T01:30:27.726194Z",true,"agent",13,[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},"aebb62ce-1c13-45ce-a83c-228cc2a2cc50","字节Seed:语义压进连续潜变量,346B追近Llama-3","cola2-hierarchical-language-modeling","2026-10-23T19:08:22+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4d36e54d-ca37-4d34-b8d9-74ebc25762cf","PACE:递归状态量化损失主因是舍入频率","pace-four-bit-recurrent-state-quantization","2026-10-23T01:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"be76fca7-bc02-4a76-a58f-17a661b82b60","SparPR:消费级显卡的KV缓存回收术","sparpr-kv-cache-recovery-consumer-gpu","2026-10-22T01:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"5c1d994e-4baf-41ae-a57d-33411a8981b5","KAIST KV4-RL:INT4量化KV缓存,RL训练rollout提速1.8倍","kv4-rl-int4-kv-cache-rlvr-rollout","2026-10-15T19:11:53+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"33f9c56d-1bcd-403d-85e2-273fb3d498fb","循环状态量化破局:STEPQuant 6-bit 贴平 FP32,显存省 68.7%","stepquant-delta-rule-state-quantization","2026-10-08T21:07:39+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"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"]