[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-slim-rl-diffusion-rl-risk":3,"news-related-9d376eae-46cf-48df-acd5-f19994948428":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},"9d376eae-46cf-48df-acd5-f19994948428","SLIM-RL:扩散大模型 RL 训练从「轨迹重构」走向「风险控制」","扩散大语言模型(dLLM)的 RL 训练长期为「随机掩码目标」与「推理轨迹」的不匹配买单——目前 SoTA TraceRL 把每个 rollout 切成 K\u002Fs 个轨迹对齐样本,训练成本随块大小 K 线性增长。arXiv:2607.00208 提出的 SLIM-RL 给出一条不同路径:完全不重构轨迹,而是用 τ-预算解码器为每一步 rollout 设定 commit risk 上限,从数据源头压住不匹配的损失。配套的 trace-free 随机掩码目标把序列级重要性采样和 masking 水平的确定性 quadrature 结合起来,再加一个「逐块均分保持、单调下降」的 mask schedule,有效降低方差。结果在 SDAR-4B 上以 0.46× 训练样本复现 TraceRL 最佳 MATH500;在匹配动态采样下,MATH500 提升 6.32%、GSM8K 提升 11.05%。更关键的是规模倒挂:经 SLIM-RL 训练的 4B 模型直接超过更大的 LLaDA-8B(MATH500 +10.76%)和 Dream-7B,但仍略低于自回归基线 Qwen2.5-7B——说明 dLLM 的 RL 收益远未触顶。同时 τ-预算解码器无需再训练即可在 LLaDA、Dream、SDAR 间迁移,意味着同一套 RL 配方可在多个开源 dLLM 上复用。代码已开源(github.com\u002Flaolaorkkkkk\u002FSLIM-RL)。这给行业一个清晰信号:扩散 LLM 的 RL 未来未必是把 rollout 切得更细,而是把「风险预算」和「随机掩码」耦合起来——一种更便宜、更可迁移的训练范式正在替代轨迹工程。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.00208","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},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"7a136518-ed76-4970-a007-c580cd1a11db","en","SLIM-RL: diffusion RL training moves to risk control","Diffusion LLM (dLLM) RL training has long been paying for the mismatch between \"random masking target\" and \"inference trajectory\" — the current SoTA TraceRL cuts each rollout into K\u002Fs trajectory-aligned samples, with training cost growing linearly with block size K. SLIM-RL, proposed in arXiv:2607.00208, takes a different path: rather than reconstructing trajectories at all, it uses a τ-budget decoder to set a commit-risk upper bound for each step's rollout, suppressing mismatched loss at the data source. The companion trace-free random-masking objective combines sequence-level importance sampling with deterministic quadrature on masking levels, plus a \"per-block uniform hold, monotonically decreasing\" mask schedule, effectively reducing variance. Results on SDAR-4B with 0.46× training samples reproduce TraceRL's best MATH500; under matched dynamic sampling, MATH500 improves 6.32%, GSM8K improves 11.05%. More critically, the scale inversion: the 4B model trained with SLIM-RL directly surpasses the larger LLaDA-8B (MATH500 +10.76%) and Dream-7B, but still slightly below the autoregressive baseline Qwen2.5-7B — indicating the RL dividend of dLLM is far from tapped. At the same time, the τ-budget decoder transfers across LLaDA, Dream, and SDAR without retraining, meaning the same RL recipe can be reused across multiple open-source dLLMs. Code is open-sourced (github.com\u002Flaolaorkkkkk\u002FSLIM-RL). This gives the industry a clear signal: the future of diffusion LLM RL may not be cutting rollouts finer, but coupling \"risk budget\" with \"random masking\" — a cheaper, more transferable training paradigm is replacing trajectory engineering.","slim-rl-diffusion-rl-risk","2026-07-05T20:30:00Z","2026-07-05T20:09:54.033119Z","2026-08-19T02:08:40.142862Z",true,"agent",104,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"cac485ab-a429-4ebc-88c6-a1f924f978ff","AWS 开源 KeysAndValues:微调时就让模型学会“遗忘”,单张 A100 撑住 128K","aws-keysvalues-sparse-attention-finetuning","2026-08-26T05:20:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"c94bdf86-5de9-49fe-8c98-0f5c47611bfe","SGLang v0.5.18 发布:大模型冷启动提速 2.38 倍,710 个 PR 都改了什么","sglang-v0-5-18-cold-start-2-38x","2026-08-24T23:15:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"1311adb6-dc19-41a7-a188-6760d9e53672","HF Summer 2026 报告:13 个下载量 Top 25 模型是 2022 年的老面孔","hugging-face-summer-2026-attention-adoption","2026-08-24T08:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"988bbfb8-672a-4c6c-98f7-3a170b6bd8b3","Macaw 把 LFM2.5 装进 1.5GB:4-bit 端侧 LLM 跑 Mac 控制工具链","macaw-lfm25-15gb-edge-mac-agent","2026-08-24T06:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"ce70384a-990b-4994-bfb6-27775be45661","TensorRT Edge-LLM 0.10.0：边端第一个统一的 C++ 多模态推理栈","tensorrt-edge-llm-0-10-multimodal-runtime","2026-08-23T00:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"92433e6b-113a-4ada-af77-fbb8995a9850","LFM2.5-DSpark 开源:300M 草稿模型让端侧推理快 2.87 倍,输出零损耗","lfm2-5-dspark-draft-models","2026-08-21T21:10:00+00:00"]