{"provider_name":"Hatena Blog","version":"1.0","type":"rich","categories":["LLM"],"provider_url":"https://hatena.blog","blog_url":"https://tech-blog.abeja.asia/","blog_title":"ABEJA Tech Blog","author_name":"go5paopao","width":"100%","height":"190","html":"<iframe src=\"https://hatenablog-parts.com/embed?url=https%3A%2F%2Ftech-blog.abeja.asia%2Fentry%2Fllm-rl-rollout-correction-202606\" title=\"On-policy \u306e\u306f\u305a\u304c Off-policy \u306b\u306a\u308b\uff1aLLM \u5f37\u5316\u5b66\u7fd2 \u306e rollout mismatch\u3068\u5bfe\u7b56(rollout correction) - ABEJA Tech Blog\" class=\"embed-card embed-blogcard\" scrolling=\"no\" frameborder=\"0\" style=\"display: block; width: 100%; height: 190px; max-width: 500px; margin: 10px 0px;\"></iframe>","image_url":"https://cdn-ak.f.st-hatena.com/images/fotolife/g/go5paopao/20260603/20260603004017.png","description":"ABEJA\u3067\u30c7\u30fc\u30bf\u30b5\u30a4\u30a8\u30f3\u30c6\u30a3\u30b9\u30c8\u3092\u3057\u3066\u3044\u308b\u670d\u90e8\u3067\u3059\u3002 \u4eca\u56de\u306fGRPO\u306a\u3069\u306eLLM\u306e\u5f37\u5316\u5b66\u7fd2\u306e\u4e2d\u3067\u5b9f\u969b\u306b\u8d77\u3053\u308a\u3046\u308b\u5c11\u3057\u30de\u30a4\u30ca\u30fc\u3060\u3051\u3069\u91cd\u8981\u304b\u3082\u3057\u308c\u306a\u3044\u5f37\u5316\u5b66\u7fd2\u306eRollout\u6642\u306b\u767a\u751f\u3059\u308b\u30c8\u30d4\u30c3\u30af\u306b\u3064\u3044\u3066\u306e\u8a18\u4e8b\u3067\u3059\u3002 \u904e\u53bb\u306b\u3082AgenticRL\u306b\u95a2\u3057\u3066\u306e\u8a18\u4e8b\u3082\u66f8\u3044\u3066\u307e\u3059\u306e\u3067\u3001\u3053\u3061\u3089\u3082\u8208\u5473\u3042\u308b\u65b9\u3044\u307e\u3057\u305f\u3089\u8aad\u3093\u3067\u3044\u305f\u3060\u3051\u308b\u3068\u5e78\u3044\u3067\u3059\u3002 tech-blog.abeja.asia \u672c\u8a18\u4e8b\u306f\u4f55\u3068\u306a\u304f\u3067\u3082GRPO\u306a\u3069\u306e\u624b\u6cd5\u3067\u4f55\u3092\u3057\u3066\u3044\u308b\u304b\u77e5\u3063\u3066\u3044\u308b\u524d\u63d0\u306e\u8a18\u4e8b\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002 \u4e0a\u8a18\u8a18\u4e8b\u3092\u30d9\u30fc\u30b9\u306b\u3057\u305f\u5185\u5bb9\u3092W&B\u69d8\u306e\u30db\u30ef\u30a4\u30c8\u30da\u30fc\u30d1\u30fc\u306b\u3082\u5bc4\u7a3f\u3057\u3066\u304a\u308a\u3001\u3053\u3061\u3089\u306b\u306f\u5f37\u5316\u5b66\u7fd2\u306e\u57fa\u790e\u304b\u3089\u66f8\u304b\u308c\u3066\u3044\u308b\u306e\u3067\u3053\u3061\u3089\u3082\u3082\u3057\u826f\u3051\u308c\u3070\u2026","title":"On-policy \u306e\u306f\u305a\u304c Off-policy \u306b\u306a\u308b\uff1aLLM \u5f37\u5316\u5b66\u7fd2 \u306e rollout mismatch\u3068\u5bfe\u7b56(rollout correction)","author_url":"https://blog.hatena.ne.jp/go5paopao/","url":"https://tech-blog.abeja.asia/entry/llm-rl-rollout-correction-202606","published":"2026-06-08 14:49:34"}