{"id":4910,"date":"2026-05-17T15:00:08","date_gmt":"2026-05-17T07:00:08","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/05\/17\/sft%e5%88%ab%e6%80%a5%e7%9d%80%e6%8e%a5rl%ef%bc%81%e4%bd%a0%e7%9a%84%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%8f%af%e8%83%bd%e4%b8%80%e7%9b%b4%e5%9c%a8%e5%b8%a6%e4%bc%a4\/"},"modified":"2026-05-17T15:00:08","modified_gmt":"2026-05-17T07:00:08","slug":"sft%e5%88%ab%e6%80%a5%e7%9d%80%e6%8e%a5rl%ef%bc%81%e4%bd%a0%e7%9a%84%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%8f%af%e8%83%bd%e4%b8%80%e7%9b%b4%e5%9c%a8%e5%b8%a6%e4%bc%a4","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/05\/17\/sft%e5%88%ab%e6%80%a5%e7%9d%80%e6%8e%a5rl%ef%bc%81%e4%bd%a0%e7%9a%84%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%8f%af%e8%83%bd%e4%b8%80%e7%9b%b4%e5%9c%a8%e5%b8%a6%e4%bc%a4\/","title":{"rendered":"SFT\u522b\u6025\u7740\u63a5RL\uff01\u4f60\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u53ef\u80fd\u4e00\u76f4\u5728\u201c\u5e26\u4f24\u8bad\u7ec3\u201d"},"content":{"rendered":"<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/f94aa2c60f7306acea0c27c0c29b31a2.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/f94aa2c60f7306acea0c27c0c29b31a2.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"SFT\u522b\u6025\u7740\u63a5RL\uff01\u4f60\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u53ef\u80fd\u4e00\u76f4\u5728\u201c\u5e26\u4f24\u8bad\u7ec3\u201d\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/d193bd0595805b1c868ee6635c03cd43.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/d193bd0595805b1c868ee6635c03cd43.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"SFT\u522b\u6025\u7740\u63a5RL\uff01\u4f60\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u53ef\u80fd\u4e00\u76f4\u5728\u201c\u5e26\u4f24\u8bad\u7ec3\u201d\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/055eaa1aed5ac513ad7b7786aa2138c3.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/055eaa1aed5ac513ad7b7786aa2138c3.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"SFT\u522b\u6025\u7740\u63a5RL\uff01\u4f60\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u53ef\u80fd\u4e00\u76f4\u5728\u201c\u5e26\u4f24\u8bad\u7ec3\u201d\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/2faec1fffc569d3675aef458d19b3dc3.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/05\/2faec1fffc569d3675aef458d19b3dc3.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"SFT\u522b\u6025\u7740\u63a5RL\uff01\u4f60\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u53ef\u80fd\u4e00\u76f4\u5728\u201c\u5e26\u4f24\u8bad\u7ec3\u201d\"\/><\/div><\/figure>\n<p>PRISM\u56e2\u961f \u6295\u7a3f<\/p>\n<p>SFT\u4e4b\u540e\uff0c\u76f4\u63a5\u4e0a\u5f3a\u5316\u5b66\u4e60\u5c31\u591f\u4e86\u5417\uff1f<\/p>\n<p>\u5c0f\u5fc3\uff0c\u4f60\u505a\u7684\u53ef\u80fd\u4e0d\u662f\u201c\u8bad\u7ec3\u201d\uff0c\u800c\u662f\u201c\u8fd8\u503a\u201d\u3002<\/p>\n<p>\u5728\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u7684\u540e\u8bad\u7ec3\u4e2d\uff0c\u884c\u4e1a\u5185\u957f\u671f\u9075\u5faa\u7740\u4e00\u4e2a\u770b\u4f3c\u5929\u7ecf\u5730\u4e49\u7684\u8303\u5f0f\uff1a\u5148SFT\uff0c\u518dRL\uff0c\u4e24\u6b65\u5230\u4f4d\u3002<\/p>\n<p>\u4eceDeepSeek\u5230Qwen\uff0c\u4eceGRPO\u5230DAPO\uff0c\u5927\u5bb6\u62fc\u547d\u4f18\u5316RL\u7b97\u6cd5\u7684\u7a33\u5b9a\u6027\u3001\u91c7\u6837\u6548\u7387\u3001\u5956\u52b1\u8bbe\u8ba1\u2026\u2026\u5374\u51e0\u4e4e\u6ca1\u4eba\u56de\u5934\u770b\u4e00\u773c\uff1a<\/p>\n<p>SFT\u5230RL\u4e4b\u95f4\uff0c\u662f\u4e0d\u662f\u5c11\u4e86\u70b9\u4ec0\u4e48\uff1f<\/p>\n<p>\u4f46\u6765\u81ea\u9999\u6e2f\u79d1\u6280\u5927\u5b66\uff08\u5e7f\u5dde\uff09\u3001\u5357\u6d0b\u7406\u5de5\u5927\u5b66\u3001\u6e05\u534e\u5927\u5b66\u7b49\u673a\u6784\u7684\u6700\u65b0\u7814\u7a76Beyond SFT-to-RL(PRISM)\u7ed9\u51fa\u4e86\u4e00\u4e2a\u4ee4\u4eba\u4e0d\u5b89\u7684\u53d1\u73b0\uff1a<\/p>\n<p>SFT\u4e0d\u4ec5\u6ca1\u6709\u4e3aRL\u94fa\u597d\u8def\uff0c\u53cd\u800c\u5728\u6084\u6084\u6316\u5751\u3002<\/p>\n<p>\u5148\u770b\u4e00\u7ec4\u6bd4\u8f83\u6709\u610f\u601d\u7684\u6570\u636e\uff087\u4e2a\u4e3b\u6d41\u591a\u6a21\u6001benchmark\u7684\u5e73\u5747\u51c6\u786e\u7387\uff09\uff1a<\/p>\n<p>\u9636\u6bb5<\/p>\n<p>Qwen3-VL-4B<\/p>\n<p>Qwen3-VL-8B<\/p>\n<p>\u539f\u59cbInstruct\u6a21\u578b<\/p>\n<p>59.7%<\/p>\n<p>63.3%<\/p>\n<p>SFT\u4e4b\u540e<\/p>\n<p>56.8% (-3.0)<\/p>\n<p>58.1% (-5.2)<\/p>\n<p>SFT \u2192 GRPO<\/p>\n<p>61.8%<\/p>\n<p>63.3%<\/p>\n<p>\u53ef\u4ee5\u770b\u5230\uff0cSFT\u4e4b\u540e\uff0c\u6a21\u578b\u6027\u80fd\u53cd\u800c\u4e0b\u964d\u4e86\u3002<\/p>\n<p>8B \u6a21\u578b\u8981\u66f4\u4e3a\u660e\u663e\u4e00\u70b9\uff1aSFT\u6389\u4e865.2\u4e2a\u70b9\uff0c\u8f9b\u8f9b\u82e6\u82e6\u505a\u5b8c\u5f3a\u5316\u5b66\u4e60\uff0c\u624d\u521a\u521a\u722c\u56de\u57fa\u7ebf(baseline)\u7684\u6c34\u5e73\uff0863.3%\u219258.1%\u219263.3%\uff09\u3002<\/p>\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c\u4f60\u7684RL\u53ef\u80fd\u4e00\u76f4\u5728\u201c\u8fd8\u503a\u201d\uff0c\u800c\u4e0d\u662f\u5728\u201c\u63d0\u5347\u201d\u3002<\/p>\n<p>\u800c\u4e14\u8fd9\u7edd\u4e0d\u662f\u4e2a\u4f8b\u3002<\/p>\n<p>\u5728\u5f53\u4e0b\u4e3b\u6d41\u7684\u5f3aInstruct\u6a21\u578b\u4e0a\uff08Qwen3-VL\u7b49\uff09\uff0c\u53ea\u8981SFT\u6570\u636e\u5e26\u5165\u4e00\u4e2a\u4e0e\u57fa\u5ea7\u4e0d\u4e00\u81f4\u7684\u65b0\u5206\u5e03\uff08\u6bd4\u5982\u76ee\u524d\u6700\u5e38\u89c1\u7684GPT\/Gemini\u84b8\u998f\u6570\u636e\uff09\u51e0\u4e4e\u90fd\u4f1a\u89c2\u5bdf\u5230\u7c7b\u4f3c\u7684\u6389\u70b9\u3002<\/p>\n<p>\u539f\u56e0\u5f88\u76f4\u63a5\uff1a\u8fd9\u7c7b\u57fa\u5ea7\u5df2\u7ecf\u7ecf\u8fc7\u5927\u89c4\u6a21\u3001\u7cbe\u7ec6\u7684\u540e\u8bad\u7ec3\uff0c\u80fd\u529b\u672c\u5c31\u5904\u4e8e\u4e00\u4e2a\u76f8\u5bf9\u7a33\u5b9a\u7684\u9ad8\u4f4d\u3002<\/p>\n<p>SFT\u903c\u7740\u6a21\u578b\u53bb\u6a21\u4eff\u4e00\u5957\u65b0\u5206\u5e03\uff0c\u7ed3\u679c\u5c31\u662f\u7528\u4e00\u4e2a\u66f4\u201c\u7a84\u201d\u7684\u5206\u5e03\u53bb\u8986\u76d6\u4e00\u4e2a\u66f4\u201c\u5e7f\u201d\u7684\u80fd\u529b\uff0c\u65e7\u80fd\u529b\u88ab\u51b2\u6389\u3001\u65b0\u80fd\u529b\u53c8\u6ca1\u771f\u6b63\u5b66\u5230\u3002<\/p>\n<p>\u6362\u53e5\u8bdd\u8bf4\uff0c\u6a21\u578b\u8d8a\u5f3a\u3001\u8d8a\u63a5\u8fd1\u5b9e\u9645\u90e8\u7f72\u7684\u6c34\u5e73\uff0cSFT\u5f15\u5165\u7684\u5206\u5e03\u504f\u79fb\u5c31\u8d8a\u6210\u4e3aRL\u4e4b\u524d\u4e00\u9053\u7ed5\u4e0d\u5f00\u7684\u201c\u6697\u5751\u201d\u3002<\/p>\n<p>\u8fd9\u6070\u6070\u662fPRISM\u5fc5\u987b\u5b58\u5728\u7684\u7406\u7531\u3002<\/p>\n<p>\u8fd9\u80cc\u540e\u7684\u6838\u5fc3\u95ee\u9898\uff0c\u662f\u540e\u8bad\u7ec3\u91cc\u65e9\u5df2\u88ab\u53cd\u590d\u8ba8\u8bba\u7684\u5206\u5e03\u6f02\u79fb\uff08Distributional Drift\uff09\u3002<\/p>\n<p>\u4f46\u5728\u591a\u6a21\u6001\u573a\u666f\u4e0b\uff0c\u5b83\u6709\u4e00\u5957\u66f4\u9690\u853d\u3001\u4e5f\u66f4\u96be\u6cbb\u7684\u8868\u73b0\u5f62\u5f0f\u3002<\/p>\n<p>SFT\u5728\u591a\u6a21\u6001\u573a\u666f\u4e0b\uff0c\u4f1a\u5f15\u5165\u4e24\u7c7b\u5bb9\u6613\u88ab\u5ffd\u89c6\u7684\u504f\u5dee\uff1a<\/p>\n<p>SFT\u7684\u4f18\u5316\u76ee\u6807\u662f\u5728\u5747\u5300\u7684token\u7ea7loss\u4e0b\u6a21\u4eff\u6f14\u793a\u8f68\u8ff9\u3002<\/p>\n<p>\u5b83\u4e0d\u533a\u5206\u201c\u8fc7\u7a0b\u201d\u548c\u201c\u7ed3\u679c\u201d\uff1a\u5bf9\u6a21\u578b\u6765\u8bf4\uff0c\u6b63\u786e\u7684\u63a8\u7406\u6b65\u9aa4\u548c\u683c\u5f0f\u5316\u7684\u6a21\u677f\u5957\u8bdd\uff0c\u6743\u91cd\u662f\u4e00\u6837\u7684\u3002<\/p>\n<p>\u7ed3\u679c\u5c31\u662f\u6a21\u578b\u5b66\u4f1a\u4e86\u201c\u957f\u5f97\u50cf\u201d\u6b63\u786e\u7b54\u6848\uff0c\u800c\u4e0d\u662f\u201c\u60f3\u5f97\u51fa\u201d\u6b63\u786e\u7b54\u6848\u3002 \u5b83\u5b66\u5230\u7684\u662f\u8868\u9762\u6a21\u5f0f\uff0c\u800c\u975e\u5fe0\u5b9e\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p>\u8fd9\u662f\u591a\u6a21\u6001\u573a\u666f\u7279\u6709\u7684\u9ebb\u70e6\u3002\u4e0e\u7eaf\u6587\u672c\u6a21\u578b\u4e0d\u540c\uff0c\u591a\u6a21\u6001\u6a21\u578b\u7684\u6f02\u79fb\u4e0d\u662f\u5355\u4e00\u7684\uff0c\u800c\u662f\u4e24\u79cd\u5b9a\u6027\u4e0d\u540c\u7684\u5931\u8d25\u6a21\u5f0f\u5728\u540c\u65f6\u53d1\u751f\uff1a<\/p>\n<p>\u8fd9\u4e24\u79cd\u6f02\u79fb\u7684\u6210\u56e0\u4e0d\u540c\u3001\u7ea0\u6b63\u65b9\u5f0f\u4e0d\u540c\uff0c\u4f46SFT\u7528\u540c\u4e00\u4e2atoken loss\u628a\u5b83\u4eec\u4e00\u8d77\u62df\u5408\u3002<\/p>\n<p>\u800c\u5f53RL\u9636\u6bb5\u65f6\uff0c\u6a21\u578b\u5df2\u7ecf\u5728\u611f\u77e5\u548c\u63a8\u7406\u4e24\u7aef\u540c\u65f6\u504f\u79fb\uff0c\u5373\u4e00\u4e2a\u201c\u65e2\u770b\u4e0d\u51c6\u3001\u53c8\u60f3\u4e0d\u5bf9\u201d\u7684\u6a21\u578b\u3002<\/p>\n<p>\u4eceGRPO\uff0c\u5230DAPO\uff0c\u518d\u5230GSPO\uff0cRL\u7b97\u6cd5\u8fd9\u4e00\u6bb5\u65f6\u95f4\u786e\u5b9e\u4e00\u76f4\u5728\u8fdb\u6b65\u3002<\/p>\n<p>\u4f46\u5b83\u4eec\u89e3\u51b3\u7684\u662fRL\u9636\u6bb5\u5185\u90e8\u7684\u95ee\u9898\uff1a\u91c7\u6837\u6548\u7387\u3001\u68af\u5ea6\u65b9\u5dee\u3001\u7b56\u7565\u5d29\u6e83\u3002\u6ca1\u6709\u4efb\u4f55\u4e00\u4e2aRL\u7b97\u6cd5\u56de\u5934\u53bb\u4fee\u590dSFT\u7559\u4e0b\u7684\u5206\u5e03\u504f\u5dee\u3002<\/p>\n<p>\u8fd9\u91cc\u4e3e\u4e2a\u4e0d\u592a\u6070\u5f53\u7684\u4f8b\u5b50\uff1a\u8fd9\u91cc\u5c31\u597d\u6bd4\u4f60\u53c2\u52a0\u767e\u7c73\u77ed\u8dd1\uff0cSFT\u4e0d\u4ec5\u6ca1\u6709\u8ba9\u4f60\u5f80\u524d\u8d70\uff0c\u53cd\u800c\u628a\u4f60\u5411\u540e\u63a8\u4e8650\u7c73\u3002<\/p>\n<p>\u73b0\u6709\u7684RL\u7b97\u6cd5\u90fd\u5728\u7814\u7a76\u600e\u4e48\u8dd1\u5f97\u66f4\u5feb\uff0c\u4f46\u8d77\u70b9\u8fd8\u5728\u5751\u91cc\uff0c\u800cPRISM\u8981\u505a\u7684\uff0c\u5c31\u662f\u5728SFT\u548cRL\u4e4b\u95f4\u8865\u4e0a\u8fd9\u4e00\u6b65\uff0c\u4e0d\u4ec5\u628a\u4f60\u62c9\u56de\u8d77\u8dd1\u7ebf\uff0c\u8fd8\u987a\u52bf\u5f80\u524d\u63a8\u4e00\u628a\uff0c\u8ba9RL\u53ea\u7528\u8dd150\u7c73\u5c31\u80fd\u51b2\u7ebf\u3002<\/p>\n<p>PRISM\u6253\u7834\u4e86\u4f20\u7edf\u7684\u4e24\u9636\u6bb5\u8303\u5f0f\uff0c\u63d0\u51fa\u4e86SFT \u2192 \u5206\u5e03\u5bf9\u9f50 (PRISM) \u2192 RLVR\u7684\u4e09\u9636\u6bb5\u6d41\u6c34\u7ebf\u3002<\/p>\n<p>\u5173\u952e\u521b\u65b0\u5728\u4e8e\u4e2d\u95f4\u7684\u5206\u5e03\u5bf9\u9f50\u9636\u6bb5\u3002<\/p>\n<p>\u611f\u77e5\u6f02\u79fb\u548c\u63a8\u7406\u6f02\u79fb\u662f\u4e24\u7c7b\u6210\u56e0\u4e0d\u540c\u7684\u504f\u5dee\uff0c\u9700\u8981\u5206\u5f00\u5904\u7406\u3002<\/p>\n<p>PRISM\u4e3a\u6b64\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u6df7\u5408\u4e13\u5bb6\u5224\u522b\u5668\uff0c\u7531\u4e24\u4e2a\u4e13\u95e8\u5316\u7684\u4e13\u5bb6\u7ec4\u6210\uff1a<\/p>\n<p>\u6700\u7ec8\u5224\u522b\u5f97\u5206\u4e3a\u4e24\u8005\u7684\u52a0\u6743\u7ec4\u5408\uff1a<\/p>\n<p>r(x,y) = \u03b1 \u00b7 D_v(x,c) + (1-\u03b1) \u00b7 D_r(x,t)<\/p>\n<p>\u8fd9\u79cd\u8bbe\u8ba1\u7684\u597d\u5904\u662f\u63d0\u4f9b\u89e3\u8026\u7684\u7ea0\u6b63\u4fe1\u53f7\uff0c\u907f\u514d\u5c06\u4e24\u79cd\u5b8c\u5168\u4e0d\u540c\u7684\u8bef\u5dee\u6a21\u5f0f\u585e\u8fdb\u4e00\u4e2a\u6807\u91cf\u91cc\uff0c\u5bfc\u81f4\u68af\u5ea6\u4fe1\u53f7\u53d8\u5f97\u5608\u6742\u3002<\/p>\n<p>PRISM\u7684\u53e6\u4e00\u4e2a\u4f18\u96c5\u4e4b\u5904\u5728\u4e8e\uff1a\u5b83\u662f\u9ed1\u76d2\u7684\u3002<\/p>\n<p>\u5f88\u591a\u84b8\u998f\u65b9\u6cd5\u9700\u8981\u8bbf\u95ee\u6559\u5e08\u6a21\u578b\u7684logits\uff08\u5185\u90e8\u6982\u7387\u5206\u5e03\uff09\uff0c\u8fd9\u610f\u5473\u7740\u4f60\u5f97\u6709\u6559\u5e08\u6a21\u578b\u7684\u5b8c\u6574\u6743\u91cd\u3002<\/p>\n<p>\u4f46\u5728\u5b9e\u9645\u573a\u666f\u4e2d\uff0c\u6700\u5f3a\u7684\u6a21\u578b\u5f80\u5f80\u53ea\u63d0\u4f9bAPI\uff0c\u4f60\u53ea\u80fd\u770b\u5230\u8f93\u51fa\uff0c\u770b\u4e0d\u5230\u5185\u90e8\u72b6\u6001\u3002<\/p>\n<p>PRISM\u5b8c\u5168\u5728\u54cd\u5e94\u7ea7\u522b\u5de5\u4f5c\uff1a\u4ece\u5f3a\u6a21\u578b\uff08Gemini 3 Flash\uff09\u91c7\u96c6\u9ad8\u8d28\u91cf\u8f93\u51fa\u4f5c\u4e3a\u6b63\u6837\u672c\uff0c\u4ece\u5f53\u524d\u7b56\u7565\u91c7\u6837\u4f5c\u4e3a\u8d1f\u6837\u672c\uff0c\u901a\u8fc7\u5bf9\u6297\u535a\u5f08\u6765\u5bf9\u9f50\u5206\u5e03\u3002<\/p>\n<p>\u53ea\u8981\u80fd\u8c03API\uff0c\u5c31\u80fd\u7528PRISM\u3002<\/p>\n<p>\u4f20\u7edfRL\u8bad\u7ec3\u901a\u5e38\u4f1a\u52a0\u4e00\u4e2aKL\u6563\u5ea6\u7ea6\u675f\uff0c\u9632\u6b62\u7b56\u7565\u504f\u79bb\u521d\u59cb\u6a21\u578b\u592a\u8fdc\u3002\u4f46PRISM\u6709\u610f\u8bc6\u5730\u53bb\u6389\u4e86\u8fd9\u4e2a\u7ea6\u675f\u3002<\/p>\n<p>\u9053\u7406\u5f88\u7b80\u5355\uff0c\u5bf9\u9f50\u9636\u6bb5\u7684\u76ee\u7684\uff0c\u5c31\u662f\u7ea0\u6b63SFT\u5e26\u6765\u7684\u5206\u5e03\u504f\u5dee\u3002\u518d\u52a0\u4e00\u4e2a\u628a\u7b56\u7565\u62c9\u56deSFT\u5206\u5e03\u7684KL\u7ea6\u675f\uff0c\u672c\u8eab\u5c31\u548c\u8fd9\u4e2a\u76ee\u6807\u76f8\u4e92\u77db\u76fe\u3002<\/p>\n<p>\u4e0b\u56fe\u76f4\u89c2\u5730\u5c55\u793a\u4e86\u5206\u5e03\u7684\u6f14\u53d8\u8fc7\u7a0b\uff1a\u4eceBase\u5230Post-SFT\u518d\u5230Post-Alignment\uff0c\u65e0\u8bba\u662f\u63a8\u7406\u6b65\u6570\u8fd8\u662f\u89c6\u89c9\u63cf\u8ff0\u9879\u6570\u7684\u5206\u5e03\uff0c\u90fd\u5728\u9010\u6b65\u5411\u76d1\u7763\u6570\u636e\u9760\u62e2\uff1a<\/p>\n<p>\u53ef\u4ee5\u6e05\u6670\u770b\u5230\uff1aPost-SFT\uff08\u84dd\u7ebf\uff09\u4e0eSupervision\uff08\u9ed1\u7ebf\uff09\u4ecd\u6709\u660e\u663e\u504f\u5dee\uff0c\u800cPost-Alignment\uff08\u6a59\u7ebf\uff09\u5219\u5927\u5e45\u7f29\u5c0f\u4e86\u8fd9\u4e00\u5dee\u8ddd\uff0c\u4e14\u8fd9\u79cd\u6539\u8fdb\u5728Post-RLVR\uff08\u7eff\u7ebf\uff09\u9636\u6bb5\u5f97\u4ee5\u4fdd\u6301\u3002<\/p>\n<p>\u5728Qwen3-VL\u76844B\u548c8B\u4e24\u4e2a\u89c4\u6a21\u4e0a\uff0cPRISM\u642d\u914dGRPO\/DAPO\/GSPO\u4e09\u79cd\u4e3b\u6d41RL\u7b97\u6cd5\uff0c\u57284\u4e2a\u6570\u5b66\u63a8\u7406\u57fa\u51c6\uff08MathVista\u3001MathVerse\u3001MathVision\u3001WeMath\uff09\u548c3\u4e2a\u901a\u7528\u591a\u6a21\u6001\u57fa\u51c6\uff08MMMU\u3001MMMU-Pro\u3001HallusionBench\uff09\u4e0a\u5168\u9762\u9a8c\u8bc1\u4e86\u6709\u6548\u6027\u3002<\/p>\n<p>\u4e0b\u8868\u662f\u8bba\u6587Table 1\u7684\u4e3b\u7ed3\u679c\uff08\u7070\u8272\u884c\u4e3aPRISM\uff09\uff1a<\/p>\n<p>\u4ece\u4e3b\u8868\u91cc\u80fd\u8bfb\u51fa\u51e0\u4e2a\u503c\u5f97\u5c55\u5f00\u7684\u4fe1\u53f7\uff1a<\/p>\n<p>(1)\u6a21\u578b\u8d8a\u5f3a\uff0cPRISM\u7684\u589e\u76ca\u8d8a\u5927\uff1a8B\u62ff\u5230+6.0\u7684\u5e73\u5747\u63d0\u5347\uff0c4B\u4e3a+4.4\uff0c\u66f4\u5f3a\u7684\u57fa\u5ea7\u88abSFT\u201c\u4f24\u5bb3\u201d\u5f97\u66f4\u6df1\uff0c\u4e5f\u56e0\u6b64\u4ece\u5bf9\u9f50\u4e2d\u53d7\u76ca\u66f4\u591a\uff1b<\/p>\n<p>(2)PRISM\u5728\u7edd\u5927\u591a\u6570\u5b50\u57fa\u51c6\u4e0a\u62ff\u5230\u4e86\u540c\u57fa\u5ea7\u4e0b\u7684\u6700\u4f73\u5206\u6570\uff08\u8868\u4e2d\u52a0\u7c97\uff09\uff0c\u8986\u76d6\u6570\u5b66\u63a8\u7406\u4e0e\u901a\u7528\u89c6\u89c9\u7406\u89e3\u4e24\u7c7b\u4efb\u52a1\uff0c\u8fd9\u610f\u5473\u7740\u5bf9\u9f50\u5e26\u6765\u7684\u4e0d\u662f\u67d0\u4e2a\u9886\u57df\u7684\u5c40\u90e8\u589e\u76ca\uff0c\u800c\u662f\u5206\u5e03\u5c42\u9762\u7684\u5168\u5c40\u6821\u51c6\u3002<\/p>\n<p>\u4ece\u6d88\u878d\u8868\uff08\u8bba\u6587Table 2\uff09\u91cc\u80fd\u76f4\u63a5\u8bfb\u51fa\u6bcf\u4e2a\u7ec4\u4ef6\u7684\u8d21\u732e\uff1a<\/p>\n<p>(1)\u53bb\u6389SFT\u9636\u6bb5\u76f4\u63a5\u638916.8\u4e2a\u70b9\uff0c\u8bf4\u660eSFT\u4f5c\u4e3a\u201c\u51b7\u542f\u52a8\u201d\u4ecd\u4e0d\u53ef\u66ff\u4ee3\uff0cPRISM\u4e0d\u662f\u8981\u53d6\u4ee3SFT\uff0c\u800c\u662f\u4fee\u590d\u5b83\u5e26\u6765\u7684\u526f\u4f5c\u7528\uff1b<\/p>\n<p>(2)\u53bb\u6389\u5bf9\u9f50\u9636\u6bb5\u63894.4\u4e2a\u70b9\uff0c\u4e0e4B\u4e3b\u8868\u7684\u63d0\u5347\u5e45\u5ea6\u5b8c\u5168\u5bf9\u5e94\uff0c\u662f\u5206\u5e03\u5bf9\u9f50\u6548\u679c\u7684\u76f4\u63a5\u8bc1\u636e\uff1b<\/p>\n<p>(3)\u5355\u4e2a4B\u5224\u522b\u5668\u66ff\u4ee3MoE\u63893.4\uff0c\u4ec5\u6587\u672c\u5224\u522b\u5668\u63893.9\u3002<\/p>\n<p>\u540e\u8005\u5c24\u4e3a\u6709\u8da3\uff1a\u6ca1\u6709\u89c6\u89c9\u611f\u77e5\u7684\u5224\u522b\u5668\u53ea\u80fd\u6355\u6349\u8868\u9762\u6a21\u5f0f\uff08\u683c\u5f0f\u3001\u6a21\u677f\u3001\u98ce\u683c\uff09\uff0c\u5bfc\u81f4\u7b56\u7565\u5b66\u4f1a\u4e86\u201c\u9e66\u9e49\u5b66\u820c\u5f0f\u5bf9\u9f50\u201d\uff0c\u542c\u8d77\u6765\u50cf\u76d1\u7763\u6570\u636e\uff0c\u4f46\u5b9e\u9645\u4e0a\u770b\u4e0d\u5230\u6240\u63cf\u8ff0\u7684\u5185\u5bb9\u3002<\/p>\n<p>PRISM\u7684\u51fa\u73b0\uff0c\u7ed9\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u540e\u8bad\u7ec3\u8303\u5f0f\u6253\u4e0a\u4e86\u4e00\u4e2a\u201c\u8865\u4e01\u201d\uff0c\u4f46\u8fd9\u4e2a\u8865\u4e01\u53ef\u80fd\u6bd4\u4e3b\u7a0b\u5e8f\u8fd8\u91cd\u8981\u3002<\/p>\n<p>SFT \u548cRL\u4e4b\u95f4\u4e0d\u662f\u65e0\u7f1d\u8854\u63a5\uff0c\u800c\u662f\u5b58\u5728\u4e00\u9053\u88ab\u957f\u671f\u5ffd\u7565\u7684\u5206\u5e03\u65ad\u5c42\u3002RL\u7b97\u6cd5\u518d\u5f3a\uff0c\u5982\u679c\u8d77\u70b9\u5c31\u6b6a\u4e86\uff0c\u8dd1\u5f97\u8d8a\u5feb\u53ea\u4f1a\u504f\u5f97\u8d8a\u8fdc\u3002<\/p>\n<p>\u8ba9\u591a\u6a21\u6001\u5927\u6a21\u578b\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u518d\u8fdb\u4e00\u6b65\uff0c\u672a\u5fc5\u8981\u9760\u66f4\u590d\u6742\u7684RL\u7b97\u6cd5\u6216\u66f4\u591a\u8bad\u7ec3\u6570\u636e\u3002<\/p>\n<p>\u628aSFT\u548cRL\u4e4b\u95f4\u8fd9\u4e00\u6b65\u5bf9\u9f50\u8865\u4e0a\uff0c\u6a21\u578b\u81ea\u7136\u4f1a\u8dd1\u5f97\u66f4\u7a33\u3002<\/p>\n<p>Arxiv\uff1ahttps:\/\/arxiv.org\/abs\/2604.28123<\/p>\n<p>Github\uff1ahttps:\/\/github.com\/XIAO4579\/PRISM<\/p>\n<p>\u5408\u4f5c\u8be6\u8be2\uff1aswang886@connect.hkust-gz.edu.cn<\/p>\n<blockquote>\n<p><strong>\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n<p>\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n<p>\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n<p>\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>PRISM\u56e2\u961f \u6295\u7a3f<\/p>\n<p>SFT\u4e4b\u540e\uff0c\u76f4\u63a5\u4e0a\u5f3a\u5316\u5b66\u4e60\u5c31\u591f\u4e86\u5417\uff1f<\/p>\n<p>\u5c0f\u5fc3\uff0c\u4f60\u505a\u7684\u53ef\u80fd\u4e0d\u662f\u201c\u8bad\u7ec3\u201d\uff0c\u800c\u662f\u201c\u8fd8\u503a\u201d\u3002<\/p>\n<p>\u5728\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u7684\u540e\u8bad\u7ec3\u4e2d\uff0c\u884c\u4e1a\u5185\u957f\u671f\u9075\u5faa\u7740\u4e00\u4e2a\u770b\u4f3c\u5929\u7ecf\u5730\u4e49\u7684\u8303\u5f0f\uff1a\u5148SFT\uff0c\u518dRL\uff0c\u4e24\u6b65\u5230\u4f4d\u3002<\/p>\n<p>\u4eceDeepSeek\u5230Qwen\uff0c\u4eceGRPO\u5230DAPO\uff0c\u5927\u5bb6\u62fc\u547d\u4f18\u5316RL\u7b97\u6cd5\u7684\u7a33\u5b9a\u6027\u3001\u91c7\u6837\u6548\u7387\u3001\u5956\u52b1\u8bbe\u8ba1\u2026\u2026\u5374\u51e0\u4e4e\u6ca1\u4eba\u56de\u5934\u770b\u4e00\u773c\uff1a<\/p>\n<p>SFT\u5230RL\u4e4b\u95f4\uff0c\u662f\u4e0d\u662f\u5c11\u4e86\u70b9\u4ec0\u4e48\uff1f<\/p>\n<p>\u4f46\u6765\u81ea\u9999<\/p>\n","protected":false},"author":1,"featured_media":4906,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[834,835,603],"class_list":["post-4910","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-sft","tag-835","tag-603"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/4910","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=4910"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/4910\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/4906"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=4910"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=4910"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=4910"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}