[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-epic-cfg-dllm-67-5pct-speedup":3,"news-related-634ba8f7-771e-4492-aacb-549112a8c91a":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},"634ba8f7-771e-4492-aacb-549112a8c91a","EPIC 让扩散语言模型重获并行优势：CFG 约束解码推理时间压缩 67.5%","扩散语言模型（DLM）通过迭代去噪一次性更新多个 token 的方式，被视为自回归 LLM 之外最值得期待的一条路径。但当遇上代码、JSON、化学分子式这类需要严格语法约束的场景，传统 CFG 约束解码在 DLM 上却把并行优势蚕食殆尽——基线方案每步都要重新词法分析、构建并最小化 DFA、再做顺序 CFG 校验，最坏情况下比无约束解码慢上四倍。arXiv 上的 EPIC（2606.00722）把这套管线拆开重做。EPIC 的三把手术刀：第一，词法分析记忆化，把相似部分输出共享的词法单元缓存下来，避免每步从零解析。第二，抛弃 DFA，改用 Earley 风格图解析直接在词法图上判定 CFG 兼容性，省去 DFA 构建与最小化的开销，也让顺序校验器不再卡在部分掩码序列上。第三，放宽的相容子集选择——先在候选集中筛出彼此相容的 token 子集，再做精确验证，把一次并行提交多个 token 的能力还给扩散解码。三招叠加，在三个基准、四个模型上把推理时间最高压缩 67.5%，相对基线引入的额外开销最高减少 90.5%。DLM 这两年几乎都在卷如何让扩散生成更快——从 d3LLM 的蒸馏提速到 DFlash 的扩散式投机解码，核心战场都是 latency。但所有这些优化只要遇到 JSON schema、SQL、化学 SMILES 这类结构化约束，就会被 4 倍减速拉下水。EPIC 把扩散并行和结构化输出这两个长期冲突的目标重新对齐，意味着 DLM 在 Agent、自动化数据抽取、代码生成等企业级场景下，终于可以同时拿到并行解码的低延迟与语法严格性。这正是 DLM 从 demo 走向生产最需要的那一类补丁。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.00722","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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"df7ae4c5-15e2-4e44-ae8e-37548caad1d0","en","EPIC restores diffusion LM parallelism, 67.5% faster decoding","Diffusion language models (DLMs), which update multiple tokens in a single iterative denoising pass, are seen as the most promising path beyond autoregressive LLMs. But when they hit scenarios that need strict grammar constraints — code, JSON, chemical molecular formulas — traditional CFG-constrained decoding eats up the parallel advantage completely: every step in the baseline has to re-tokenize, build and minimize a DFA, then do sequential CFG checks, and in the worst case it can be 4× slower than unconstrained decoding. The EPIC paper (arXiv:2606.00722) tears this pipeline apart and rebuilds it. Three surgical cuts in EPIC: first, lexer memoization, caching the lexical units shared by similar partial outputs, avoiding parsing from scratch every step. Second, ditching the DFA in favor of Earley-style graph parsing to directly judge CFG compatibility on the lexical graph, eliminating the cost of DFA construction and minimization, and freeing the sequential checker from getting stuck on partial-mask sequences. Third, a relaxed compatibility-subset selection — first filter out mutually compatible token subsets from the candidate set, then do precise validation, returning the ability to commit multiple tokens in parallel to diffusion decoding. Combined, the three cuts cut inference time by up to 67.5% across three benchmarks and four models, and cut the extra overhead over the baseline by up to 90.5%. Over the past two years, DLMs have mostly been racing to make diffusion generation faster — from d3LLM's distillation speedup to DFlash's diffusion-style speculative decoding, the core battlefield has been latency. But every one of these optimizations hits a 4× slowdown as soon as it encounters structured constraints like JSON schema, SQL, or chemical SMILES. EPIC re-aligns two long-conflicting goals — diffusion parallelism and structured output — meaning DLMs can finally get both the low latency of parallel decoding and the grammar strictness needed for enterprise-level scenarios like agents, automated data extraction, and code generation. This is exactly the kind of patch DLM most needs to go from demo to production.","epic-cfg-dllm-67-5pct-speedup","2026-06-09T16:00:00Z","2026-06-09T16:19:04.367508Z","2026-08-19T02:08:40.142862Z",true,"agent",131,{"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},"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",{"id":59,"title":60,"news_slug":61,"published_at":62},"b45f5c46-c982-4410-9452-07a9f779218f","On-Policy Distillation 把扩散语言模型训练成本砍到 1\u002F15~1\u002F7000","opdlm-on-policy-distillation-dllm-1-7000x","2026-06-08T04:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"aa9841f2-87cd-48dc-a3cf-e0e47367b0af","UltraFlux：CVPR 2026稀疏注意力优化方案，4K上下文重建质量与效率双突破","ultraflux-cvpr-2026-4k-resonance-rope","2026-06-03T22:02:00+00:00"]