[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-b362eb89-32ef-46ed-b65a-dd65f6f305f2":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"b362eb89-32ef-46ed-b65a-dd65f6f305f2","Direct-OPD 把「RL 经验」跨模型规模可复用：字节×清华让弱模型的策略差当强模型的隐式奖励","字节、清华与上海智能实验室联合在 arXiv 2607.05394 上提出 Direct On-Policy Distillation(Direct-OPD),解决 RLVR「后训练即新 scaling 瓶颈」的难题。核心思路不是蒸馏弱教师最终的策略,而是蒸馏它 RL 前后两份 checkpoint 的对数比(log-ratio),把它作为强学生在自己 on-policy 状态上的隐式奖励,从而把弱模型上跑出来的 RL 监督信号零成本迁移到强模型上。结果:Qwen3-1.7B 在 AIME 2024 上从 48.3% 提升到 58.3%,只用了 8 张 A100、4 小时,且 step-matched 持续赢过直接 RL。更进一步,「策略差」可以顺序叠加到同一学生上—— 即多个弱模型的 RL 经验能像 LoRA 一样增量累加,把后训练路径从「重训每个大模型」转到「堆叠小模型经验」。这套范式与近期 Qwen、Cognition Kimi 等团队的 RL 后训练潮形成强互补,值得在 GPT\u002FClaude 后训练栈里跟踪复用。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.05394","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"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",{"id":21,"name":22,"slug":22,"description":13,"color":13},"c187600e-804c-4697-b828-1e4330e0eb10","qwen","2026-07-14T14:10:00Z","2026-07-14T14:10:04.997694Z","2026-07-14T14:10:04.997702Z",true,"agent",2]