[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-seer-dmllm-padding-31x":3,"news-related-5ada7ebb-5e6d-485b-9ca9-ce3d6f97b558":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},"5ada7ebb-5e6d-485b-9ca9-ce3d6f97b558","Seer 把 DMLLM 的「废 padding」一次砍掉 31× 吞吐：首个去噪第 0 步就能定位语义边界的训练免费加速框架","扩散多模态大模型(DMLLM)在多模态推理上表现出色,但固定长度生成机制一直拖累推理效率——模型必须把输出 padding 到预设的最大长度,大量无用 [EOS] 位置会被反复去噪,造成严重冗余。Zhao、Sun 和 Yan 提出的 Seer 框架发现一个关键现象:DMLLM 在第一次去噪步骤中,MLP 激活的稀疏度会发生明显跳变,这就是有效语义边界的位置信号。基于此,Seer 用信噪比(SNR)判据在 Step-0 一刀切掉所有冗余后缀,再在后续每个去噪步都跳过这部分计算,实现训练免费的「零再训练」加速。批量服务时,Seer 配合混合执行策略兼顾动态序列长度,在 9 个基准测试中保持性能不退化,部分任务(如 DocVQA 63.52→63.66)反而因为抑制了 [EOS] 位置的噪声泄漏而略有涨点,最高可带来 ~31× 吞吐加速,作为 ACM Multimedia 2026 接收工作提供了一套即插即用的 DMLLM 推理加速方案。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.14557","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"5969eb41-a732-4068-be7e-d6d3f55348e9","en","Seer cuts DMLLM wasted padding for 31x throughput","Diffusion Multimodal Large Models (DMLLM) perform well on multimodal reasoning, but their fixed-length generation mechanism has long dragged down inference efficiency — the model must pad its output to a preset maximum length, and a large number of useless [EOS] positions are repeatedly denoised, causing serious redundancy. The Seer framework proposed by Zhao, Sun and Yan discovers a key phenomenon: in the first denoising step of a DMLLM, the sparsity of the MLP activations changes dramatically, and this is the position signal of the effective semantic boundary. Based on this, Seer uses a Signal-to-Noise Ratio (SNR) criterion to cut off all redundant suffixes at Step-0 with one cut, then skips this part of the computation at every subsequent denoising step, achieving a \"zero retraining\" speedup. In batch service, Seer pairs a mixed-execution strategy with dynamic sequence length, holding performance across 9 benchmarks and even slightly improving on some tasks (e.g., DocVQA 63.52→63.66) by suppressing noise leakage from [EOS] positions, with up to ~31× throughput speedup. As work accepted by ACM Multimedia 2026, it provides a drop-in DMLLM inference-acceleration solution.","seer-dmllm-padding-31x","2026-07-19T12:15:00Z","2026-07-19T12:13:49.815567Z","2026-08-19T02:08:40.142862Z",true,"agent",83,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"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":44,"title":45,"news_slug":46,"published_at":47},"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",{"id":49,"title":50,"news_slug":51,"published_at":52},"0b744f65-a3d5-47d8-84d1-eb25a7a2798e","FMLM+ 把扩散语言模型的「自纠错」解锁：32× 更少 NFE 匹配离散基线","fmlm-plus-posterior-refinement-32x-nfe","2026-06-25T02:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"844a2f3b-682d-4d03-b8ab-f02ec1c16dcf","ImageWAM 抛弃视频生成：用图像编辑做世界动作模型，FLOPs 降到 1\u002F6","imagewam-sjtu-world-action-model-1-6-flops","2026-06-22T06:20:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"491873fe-0a02-404b-a3ba-a2490b35ec7d","TimeProVe：长视频问答的「先提议后验证」架构，把 VLM 从「全局审片」改为「点穴验证」","timeprove-propose-verify-long-video-qa","2026-06-20T08:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"80bc5e25-d24d-45a1-9c2b-534ebcae39f9","腾讯 WeDLM 开源：让扩散 LLM 在标准因果注意力下跑出 3-6× vLLM 加速","tencent-wedlm-diffusion-llm-causal-3-6x","2026-06-16T20:00:00+00:00"]