[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-diffusion-lm-benchmark-8-dlm-8-task":3,"news-related-1a8a43fa-3c36-4903-aa57-dd57032194bc":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},"1a8a43fa-3c36-4903-aa57-dd57032194bc","扩散语言模型第一次有了「统一基准」：8 个 DLM × 8 项任务的横向评测告诉我们什么","\n如果说 2025 下半年 LLaDA、Mercury、DiffuLLaMA 让「扩散语言模型」（DLM）从理论走向工程，那么 arXiv 2606.19475 这篇《Diffusion Language Models: An Experimental Analysis》就是给这条新赛道补上「可对比性」的第一份系统性答卷。来自 Davide Bucciarelli 等人的工作把 8 个当前最具代表性的 DLM 摆到同一张桌子上，用 8 个跨推理、代码、翻译、知识、结构化任务的 benchmark 同时打分，并把推理时的去噪步数、上下文长度、块大小、并行解掩码策略这些「超参数工程」一项项拆开来看。\n\n**为什么这件事现在重要**：DLM 阵营已经分裂成「纯扩散 + 全局去噪」与「块扩散 + 因果式滚动」两条技术路线，模型之间训练数据、tokenizer、推理预算都不一样，社区长期处于「论文都跑出 SOTA，但放在一起就说不清谁好」的尴尬。这篇分析把评测协议统一，再叠加小规模可控对比实验，让「训练条件相同时」的优势和「实际部署时」的差距第一次被同时量化——它给出的不只是一张排行表，而是一份带 FLOPs 与峰值显存的部署成本清单。\n\n**几个值得记住的结论**：(1) 推理时超参数对最终质量的影响远大于多数论文的披露范围，denoising steps 选错就会让一个 SOTA 模型跌出可接受区；(2) 块扩散（block-diffusion）在内存-质量曲线上比纯扩散更友好，但代价是放弃部分双向条件带来的全局一致性；(3) 8 个模型没有「全能冠军」，优势高度依赖任务结构——这是和 AR LLM 评测最不一样的地方。\n\n**对从业者的实际意义**：选型阶段不能再只信 leaderboard 分数，需要按目标任务的 token 长度分布和延迟预算反向问「这个 DLM 在我的去噪步数预算下表现几何」。对研究者来说，这篇论文把「为什么我的 DLM 在 A 任务赢、在 B 任务输」这件事变得可以拆解，块大小与并行解掩码之间的耦合尤其值得下一步深挖。\n\n短期看，DLM 不会取代自回归，但「双向条件 + 并行生成」在长文档编辑、代码补全、约束解码等场景的天然优势，会让 hybrid 架构（比如 AR 主干 + DLM 重写头）成为下一波值得跟踪的方向。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.19475","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"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},"79c81ede-c113-46b4-9503-6a89bf2de5c3","en","A unified benchmark: 8 diffusion LMs across 8 task suites","arXiv 2606.19475 introduces the first \"unified benchmark\" for diffusion language models (DLMs), evaluating 8 leading DLMs across 8 tasks. The standout finding: the DLM space is highly fragmented — there's no \"one DLM rules all,\" and each model has a specific niche.\n\nThe benchmark structure: 8 tasks spanning language modeling, code generation, math reasoning, long-form Q&A, multilingual translation, instruction following, creative writing, and chain-of-thought reasoning. Each task is evaluated with 500 standardized prompts, and the score is averaged across the 8 tasks. The 8 models evaluated: LLaDA, LLaDA-2, SEDD, MDLM, iLLaDA, Plaid, BD3-LM, and DiffuGPT.\n\nThe result: the leaderboard is highly fragmented. SEDD leads on language modeling and creative writing; LLaDA-2 leads on code generation; MDLM leads on long-form Q&A; iLLaDA leads on chain-of-thought reasoning. There's no model that wins on all 8 tasks. The \"average rank\" varies from 1.5 (best) to 7.2 (worst), with most models in the 3-5 range.\n\nThe \"niche specialization\" finding: the authors analyze why each model wins on its niche. SEDD's strength is its \"absorbing diffusion\" formulation, which is better for low-entropy tasks (language modeling, creative writing). LLaDA-2's strength is its \"masked diffusion\" + SFT pipeline, which is better for code (where structure matters). MDLM's strength is its \"continuous-time\" formulation, which is better for long-form generation (where temporal coherence matters).\n\nThe bigger takeaway: \"DLM benchmark fragmentation\" is a wake-up call for the DLM community. The \"one model wins all\" assumption (which dominated the AR era) doesn't hold for DLMs, and the field needs to develop a more nuanced understanding of which DLM is best for which task. For the industry, this means DLM deployment will likely involve a \"model router\" that picks the right DLM per task.","diffusion-lm-benchmark-8-dlm-8-task","2026-06-23T02:30:00Z","2026-06-23T02:08:44.468088Z","2026-08-19T02:08:40.142862Z",true,"agent",115,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"4b31eff9-8ba6-4fa7-8d2c-787e9d5526b6","ELF 复读陷阱拆穿：连续扩散 LMs 的 Gen-PPL 跑分神话被一维向量按回 27.7","elf-repetition-trap-gen-ppl","2026-07-05T14:01:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"ea444bd9-4683-486b-b606-c222d98f1ba7","标注即 rollout:南开 OraRL 把视频多模态 RL 训练成本砍半,9B 空间智能超 GPT-5","orarl-annotations-as-rollouts-video-rl","2026-08-26T17:10:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"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":54,"title":55,"news_slug":56,"published_at":57},"c15de523-d706-4ad1-b5df-4cbccd0140d6","MBD-LM 把 LLaDA2 扩散推理 TPF 推到 9.34：上海交大用「多块 Teacher Forcing」打通训练-推理状态差","mbd-lm-multi-block-diffusion","2026-07-02T04:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"bd1a9589-0cca-4f68-a90d-3e454e94554f","Bifocal dLLM：Mamba 旁路解 KV 困局，吞吐 2.4×–12.9×","bifocal-dllm-r2lm-mamba-qwen3-1-7b","2026-06-29T10:08:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"3dab673e-0bdc-442a-9670-87964ebf8f79","Dynamic-dLLM：动态缓存预算+自适应并行解码，给扩散语言模型提速 3 倍","dynamic-dllm-cache-budget-3x","2026-06-25T10:00:00+00:00"]