{"id":7072,"date":"2025-07-08T06:06:00","date_gmt":"2025-07-07T21:06:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7072"},"modified":"2025-07-06T11:13:43","modified_gmt":"2025-07-06T02:13:43","slug":"spiral-self-play-on-zero-sum-games-incentivizes-reasoning-via-multi-agent-multi-turn-reinforcement-learning","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7072","title":{"rendered":"SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning\u00a0<\/strong>[27.2]<br>SPIRAL\u306f\u3001\u30e2\u30c7\u30eb\u3092\u30de\u30eb\u30c1\u30bf\u30fc\u30f3\u3001\u30bc\u30ed\u30b5\u30e0\u30b2\u30fc\u30e0\u3067\u5b66\u7fd2\u3057\u3001\u81ea\u8eab\u306e\u30d0\u30fc\u30b8\u30e7\u30f3\u3092\u7d99\u7d9a\u7684\u306b\u6539\u5584\u3059\u308b\u30bb\u30eb\u30d5\u30d7\u30ec\u30a4\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3067\u3042\u308b\u3002 SPIRAL\u3092\u7528\u3044\u308b\u3053\u3068\u3067\u3001\u30bc\u30ed\u30b5\u30e0\u30b2\u30fc\u30e0\u4e0a\u3067\u306e\u30bb\u30eb\u30d5\u30d7\u30ec\u30a4\u306f\u3001\u5e83\u304f\u79fb\u52d5\u3059\u308b\u63a8\u8ad6\u80fd\u529b\u3092\u751f\u307f\u51fa\u3059\u3002 \u5206\u6790\u306b\u3088\u308a, \u3053\u306e\u4f1d\u9054\u306f, \u7cfb\u7d71\u7684\u5206\u89e3, \u671f\u5f85\u5024\u8a08\u7b97, \u30b1\u30fc\u30b9\u30fb\u30d0\u30a4\u30fb\u30b1\u30fc\u30b9\u5206\u6790\u3068\u3044\u30463\u3064\u306e\u8a8d\u77e5\u7684\u30d1\u30bf\u30fc\u30f3\u3092\u901a\u3058\u3066\u8d77\u3053\u308b\u3053\u3068\u304c\u660e\u3089\u304b\u3068\u306a\u3063\u305f\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2506.24119v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2506.24119v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Mon, 30 Jun 2025 17:58:13 GMT)<\/li>\n\n\n\n<li>\u4eba\u3078\u306e\u4f9d\u5b58\u3092\u5c11\u306a\u304f\u3059\u308b\u305f\u3081\u300cWe introduce SPIRAL, a self-play framework where models learn by playing multi-turn, zero-sum games against continuously improving versions of themselves, eliminating the need for human supervision.\u300d\u3068\u3044\u3046\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3092\u63d0\u6848\u3001\u52b9\u679c\u3092\u78ba\u8a8d\u3068\u306e\u3053\u3068\u3002\u300cKey Findings. Training on zero-sum games produces reasoning capabilities that transfer broadly.\u300d\u3068\u3057\u3066\u3044\u308b\u3002\u300cOur empirical results show that training on Kuhn Poker alone improves mathematical reasoning by 8.7% average and Minerva Math by 18.1%, surpassing models trained on 25,000 expert demonstrations\u300d\u3068SFT\u3092\u4e0a\u56de\u3063\u3066\u3044\u308b\u306e\u306f\u82e5\u5e72\u9a5a\u304d\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/spiral-rl\/spiral\">GitHub &#8211; spiral-rl\/spiral: SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[356],"class_list":["post-7072","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-self-x"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7072","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7072"}],"version-history":[{"count":3,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7072\/revisions"}],"predecessor-version":[{"id":7075,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7072\/revisions\/7075"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7072"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7072"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7072"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}