[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-alodlm-adaptive-looped-diffusion":3,"topics-all":38,"news-related-a78b61ba-6899-4b4f-86b0-8d6c1753308b":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},"a78b61ba-6899-4b4f-86b0-8d6c1753308b","亚马逊 ALoDLM:扩散语言模型学会按难度分配算力","Amazon AGI 团队把 Qwen3 骨干改造成 ALoDLM:有把握的 token 提前提交为上下文,没把握的保留潜变量多轮精炼,计算按难度分配。8B 版 11 项基准平均 80.3 超自回归基线,GSM8K 吞吐约为 Qwen3-8B 的 2.7 倍。权重已开放,许可 CC BY-NC 4.0。","扩散语言模型(DLM)的并行生成快,但质量和同规模自回归模型比总有差距。Amazon AGI 团队(与 UIC、高丽大学合作)10 月 3 日挂在 arXiv 的 ALoDLM 给出新诊断——差距源于「计算量-难度错配」:部分遮蔽的序列里,有的未知 token 看一眼上下文就能猜出,有的要算很久;而现有 DLM 在每个去噪步骤里,对所有未知位置施加**完全相同的计算深度**。([arXiv:2610.04198](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.04198))\n\nALoDLM 把「一刀切」换成「token 级自适应循环」。每个去噪步骤(外循环)内部加一个内循环:置信度高的 token 提前提交,embedding 作为离散上下文喂给后续轮次;没把握的保留潜状态继续精炼,多做几轮。何时停由学出的 halting 策略决定。训练侧,token 级计算调度被建模成潜变量,条件负证据下界(NELBO)让预测和算力分配联合学习。\n\n## 跑分:8B 平均 80.3,把 AR 基线也压过去\n\n按论文 Table 1(OpenCompass 协议、贪心解码、4096 token 上限),ALoDLM-8B 在 11 项基准平均 **80.3**,高于自回归基线 Qwen3-8B 的 78.5,也高于全部 DLM 对照(WeDLM-8B 75.1)。1.7B 档平均 65.5,同样压过 Qwen3-1.7B 的 63.8。\n\n亮点项:GPQA-Diamond 8B 档 49.5 对基线 47.0;MMLU-Pro 63.9 对 56.8;HumanEval+ 84.2 对 79.3。但也有诚实弱点:1.7B 档 MATH-500 只有 61.8,落后基线 72.0 有 10 个点;8B 档 MMLU 76.6 持平、MATH-500 80.8 略低于 81.8。扩散改造成本清晰可见。\n\n## 速度:B200 上 GSM8K 吞吐约为 Qwen3-8B 的 2.7 倍\n\n论文在单块 B200 上用 vLLM + 深度感知 KV 缓存做单流推理:GSM8K 上 ALoDLM-8B 吞吐约为 vLLM 版 Qwen3-8B 的 **2.7 倍**。同精度 93.25% 对齐时,优化后 ALoDLM-8B 达 612.4 tokens\u002Fs,WeDLM-8B 为 564.2。推理旋钮:熵阈值 τ 从 0.1 拧到 0.6,吞吐 278.7 升到 508.3 tokens\u002Fs,精度 93.8% 滑到 92.3%;退出阈值 q 调深则质量升、吞吐降。\n\n## 冷水:首 token 延迟、非商用许可、零第三方验证\n\n论文 limitations 和 HF model card 都很直白。其一,循环结构要求 prefill 阶段就为全部循环深度构建 KV 缓存,首 token 延迟高于同规模 AR 模型,短回复场景收益被削弱。其二,生成速度本质输入依赖,官方明确「加速比不应期望在每个工作负载上成立」。其三,评测主要覆盖英文推理与编程,长上下文未建立。其四,对想部署的团队最关键——**权重挂在 CC BY-NC 4.0 非商用许可下**;权重刚放出,HF 月下载量还在两位数,benchmark 全部是官方自报口径。\n\n## 所以呢\n\nALoDLM 的看点不在「又一个扩散语言模型」,而在把「按难度分配计算」这个直觉用潜变量推断变成了可训练目标:简单 token 早提交、难 token 多循环,训练推理同一套门控。τ 换速度、q 换质量的旋钮给了成本敏感场景测试时计算的自由度。但 80.3 对 78.5 的领先幅度,以及「算力分配」这个角度本身,都只有官方一份答卷。值得盯:token 级循环是否搬上更大骨干,以及 WeDLM 阵营的反扑。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.04198","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},"ea708a0b-1124-468f-b2b2-b2df7cee168d","en","Amazon ALoDLM: Diffusion LMs Get Token-Adaptive Compute","ALoDLM (Amazon) commits easy tokens early, loops hard ones. 8B averages 80.3 over 11 benchmarks, beating AR; GSM8K throughput ~2.7x Qwen3-8B.","Diffusion language models (DLMs) have spent the year on the same problem: parallel generation is fast, but quality still trails comparably-sized autoregressive models. ALoDLM, posted to arXiv on October 3 by Amazon's AGI team with UIC and Korea University collaborators, offers a new diagnosis — a \"computation-difficulty mismatch\": some unknown tokens are guessable from context at a glance, others demand far more computation, yet existing DLMs apply **identical computational depth** to every unknown position at each denoising step. ([arXiv:2610.04198](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.04198))\n\nALoDLM swaps the uniform scheme for token-adaptive recurrence. Inside each denoising step (outer loop) sits an inner loop: high-confidence tokens commit early, feeding their embeddings back as discrete context; unresolved positions retain and refine their latent states through extra recurrent passes. A learned halting policy decides when to stop. For training, token-wise computation schedules become latent variables with a derived conditional negative evidence lower bound (NELBO), so prediction and compute allocation are learned jointly.\n\n## Benchmarks: 80.3 average at 8B, AR baseline included\n\nPer Table 1 (OpenCompass protocol, greedy decoding, 4096-token limit), ALoDLM-8B — converted from the Qwen3-8B backbone — averages **80.3** across 11 benchmarks, above autoregressive Qwen3-8B's 78.5 and every DLM baseline shown (WeDLM-8B 75.1, SDAR-8B 74.2). The 1.7B model averages 65.5 versus Qwen3-1.7B's 63.8.\n\nHighlights: GPQA-Diamond 49.5 at 8B versus 47.0; MMLU-Pro 63.9 versus 56.8; HumanEval+ 84.2 versus 79.3. At 1.7B, GPQA-Diamond 41.4 against 31.3. Honest weak spots remain: MATH-500 at 1.7B lands at 61.8, ten points behind the baseline's 72.0; at 8B, MMLU merely ties at 76.6 and MATH-500's 80.8 sits just below 81.8.\n\n## Speed: ~2.7x Qwen3-8B GSM8K throughput on B200\n\nOn a single NVIDIA B200 with vLLM plus depth-aware KV caching, ALoDLM-8B delivers roughly **2.7x** the throughput of vLLM-served Qwen3-8B on GSM8K at comparable or higher accuracy, while beating the diffusion baselines shown. At matched 93.25% accuracy, optimized ALoDLM-8B reaches 612.4 tokens\u002Fs against WeDLM-8B's 564.2. The two inference knobs: raising the entropy threshold from 0.1 to 0.6 lifts throughput from 278.7 to 508.3 tokens\u002Fs while accuracy slides from 93.8% to 92.3%; deepening the exit threshold trades throughput for quality.\n\n## Cold water: TTFT, non-commercial license, zero third-party validation\n\nThe paper's limitations section and the HF model card are blunt. First, the recurrent structure forces KV-cache construction for the full recurrent depth during prefill, so time-to-first-token runs longer than a comparable AR model — eroding gains for short, latency-sensitive responses. Second, generation speed is inherently input-dependent; the card warns reported speedups \"should not be expected to hold for every workload.\" Third, evaluation covers English reasoning and programming; other languages and long-context behavior are unestablished. Fourth — decisive for deployment — **the weights ship under CC BY-NC 4.0**, not the Apache\u002FMIT norm, and with monthly downloads still in the double digits, every benchmark number is the vendor's own reporting.\n\n## So what\n\nALoDLM's significance isn't \"another DLM\" — it's that \"spend compute where the difficulty is\" became a trainable objective via latent-variable inference: easy tokens commit early, hard tokens loop longer, one gating logic shared across training and inference. The tau-for-speed, q-for-quality knobs give cost-sensitive serving a test-time compute dial. But the 80.3-versus-78.5 margin, and the compute-allocation angle itself, rest on a single vendor's answer sheet. Watch whether token-level looping gets ported to larger backbones, and how the WeDLM camp responds.","alodlm-adaptive-looped-diffusion","2026-10-07T13:15:53Z","2026-10-07T13:16:25.909845Z","2026-10-07T13:16:25.909852Z",true,"agent",597,[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},"c448a366-372a-47dc-936f-8ea89103d238","Pivot-SD:200道题让扩散LLM数学涨6分","pivot-sd-diffusion-lm-self-distillation","2026-10-05T17:09:21+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"]