[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-icml-2026-flexibility-trap":3,"news-related-d6afa8e3-b342-41da-839b-840c02c42cc8":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"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 杰出论文奖两篇获奖作品同时花落扩散模型,这种巧合在 ML 三大顶会历史上屈指可数,背后更像是一种集体判断:扩散模型已经进入「纠偏」与「补基建」的深水区。\n\n清华黄高团队与 Zanlin Ni 等人的《The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models》是这届最尖锐的技术反叛。扩散大语言模型 (dLLM) 一直被宣传的核心卖点是「任意顺序生成」——区别于 GPT、Claude 这种从左到右逐 token 蹦的自回归范式,dLLM 像画画一样从噪声里去噪出完整文本,理论上可以先写中间再写开头。但论文用大量实验证明,这个「灵活性」本身就是陷阱:模型为了支持所有可能的生成顺序,反而在每种具体顺序上都做得更差。在通用推理任务上,dLLM 实际上会绕开那些高不确定性的「分叉 token」,导致解空间多样性崩溃。作者提出 JustGRPO:在 RL rollout 阶段回归最朴素的从左到右顺序,推理阶段仍保留并行解码。这个简洁方案戳破了过去两年 dLLM 文献里被反复引用的核心假设——围绕「任意顺序」投入的大量算力和工程优化,理由可能根本站不住。\n\n并列的《High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions》则从理论侧把扩散采样精度画了新天花板:用一阶拒绝采样把 ε-误差所需评分函数调用从 poly(1\u002Fε) 降到 O(d·polylog(1\u002Fε)),把「NFE 还能砍多少倍」从工程优化推到理论上界。\n\n两篇杰出一破一立,加上 DeepMind 2016 年的 A3C 拿下 Test of Time Award——ICML 的信号很明确:扩散语言模型正走出「概念验证」阶段,真正缺的不再是更多花样,而是更冷静的审视、更扎实的理论边界,以及对「灵活即优势」这种直觉的彻底清算。","https:\u002F\u002Fblog.icml.cc\u002F2026\u002F07\u002F05\u002Fannouncing-the-icml-2026-awards\u002F","6569a0a5-7899-4c1c-8ba7-3d03834e87b5",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"c1981f86-815b-4d35-a20a-7bcdd0056b03","en","ICML 2026 best paper: any-order generation is a diffusion trap","Two ICML 2026 Outstanding Paper Award winners both landed on diffusion models — this coincidence is rare in the history of the top three ML conferences, and behind it is more like a collective judgment: diffusion models have entered the deep-water zone of \"correction\" and \"infrastructure filling\". Tsinghua's Gao Huang team and Zanlin Ni et al.'s \"The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models\" is the sharpest technical rebellion of this conference. The core selling point that diffusion LLMs (dLLM) have always touted is \"arbitrary-order generation\" — unlike GPT or Claude's left-to-right token-by-token autoregressive paradigm, dLLMs denoise complete text from noise, theoretically writing the middle first then the opening. But the paper uses extensive experiments to prove that this \"flexibility\" is itself a trap: in order to support all possible generation orders, the model actually does worse on each specific order. On general reasoning tasks, dLLMs actually avoid those high-uncertainty \"branching tokens\", causing solution-space diversity to collapse. The authors propose JustGRPO: return to the most naive left-to-right order during RL rollout, while preserving parallel decoding during inference. This simple solution punctures the core assumption that has been repeatedly cited in dLLM literature over the past two years — the substantial compute and engineering optimization invested around \"arbitrary order\" may not stand at all. The co-awarded \"High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions\" draws a new ceiling on diffusion sampling accuracy from the theoretical side: using first-order rejection sampling to bring the score function calls required for ε-error down from poly(1\u002Fε) to O(d·polylog(1\u002Fε)), pushing \"how many more times can NFE be cut\" from engineering optimization to theoretical upper bound. The two outstanding papers break and establish, plus DeepMind's 2016 A3C winning the Test of Time Award — ICML's signal is clear: diffusion language models are moving past the \"proof of concept\" stage; what they really lack now is not more variety, but calmer scrutiny, more solid theoretical boundaries, and a thorough reckoning of the intuition that \"flexibility is advantage\".","icml-2026-flexibility-trap","2026-07-06T10:00:00Z","2026-07-06T10:16:06.081952Z","2026-08-19T02:08:40.142862Z",true,"agent",183,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"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":44,"title":45,"news_slug":46,"published_at":47},"43eda321-b0b7-4df7-b20e-9758cbab42c9","记忆越完整,眼前题越做不对:MemTrapBench 把 LLM 长期记忆框架打回原形","memtrapbench-llm-memory-cognitive-traps","2026-08-22T04:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"88105269-9641-44c3-a705-1cf07314614f","LLM 思维链能看出\"用了几分力\":SARE 给每一步推理做 CT 扫描","step-aware-reasoning-energy-llm-cot","2026-08-04T04:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"4b693fb4-541f-47ed-8923-6e280cec965f","大模型的“记忆”还没过视觉这一关：MEMLENS 把长上下文的短板测出来了","memlens-multimodal-long-term-memory","2026-08-03T02:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"0fd9ee7a-5b8f-49d2-9032-57f763de40e3","OpenAI 下一代模型 Astra 一口气破解 10 个数学难题:从 27 年未决的非 sofic 群到 46 年未动的高维球体堆积","openai-astra-ten-math-proofs-2026","2026-08-01T10:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"5bfdf32b-44eb-4eb5-a98b-39e921168182","九天内连发五款前沿模型:7 月的大模型军备赛,真正决胜负的不再是 benchmark","july-2026-five-frontier-models","2026-07-23T12:00:00+00:00"]