[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-5c7369f7-8635-4a6c-b5a4-f13691eecb78":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":20,"created_at":21,"modified_at":22,"is_published":23,"publish_type":24,"image_url":13,"view_count":25},"5c7369f7-8635-4a6c-b5a4-f13691eecb78","IBM Research 用 DASH 把推理模型的「过度思考」砍下来:AIME25 比 GRPO 多 5.4 个点,靠的是中间答案当免费监督","IBM Research(Chia-Hsuan Lee 等,7 月 1 日挂 arXiv:2607.00482)提出 DASH,从机制层回答推理大模型的「过度思考」问题:哪段自反思有效,哪段是无效空转。\n\n论文先打消融:即便控制长度,错误轨迹里无效自反思比例仍系统性偏高——长度不是原罪,方向才是。step-level 标注太贵,DASH 的关键洞察是:推理链里每一次「我先尝试答案 X」都是天然检查点,与 ground truth 比一比就能无监督判定后续反思是收敛还是漂移,据此把整条轨迹切成可单独打分的 segment,在优势函数里按段塑形。\n\nAIME25 上 DASH 把 GRPO 从 45.4% 抬到 50.8%,行为上少绕弯路、自纠更有效。意义不只是 +5.4,而是不引入额外监督——所有 RLVR 流水线都能低成本接入。\n\n后训练信用分配始终卡在轨迹级或 token 级。IBM 把 credit 切成 segment 级,给 Qwen、Claude、Gemini、DeepSeek 等长 CoT 推理模型提供一条通用的「降 token + 涨正确率」工程路径。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.00482","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"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},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-07-08T06:12:00Z","2026-07-07T22:13:51.814033Z","2026-07-07T22:13:51.814043Z",true,"agent",2]