
{"id":11133,"date":"2026-08-18T02:18:25","date_gmt":"2026-08-17T18:18:25","guid":{"rendered":"https:\/\/infernews.com\/blog\/mmdiff\/"},"modified":"2026-08-18T02:18:25","modified_gmt":"2026-08-17T18:18:25","slug":"mmdiff","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/mmdiff\/","title":{"rendered":"MMDiff \u4ee5\u7279\u5fb5\u5dee\u5206\u64cd\u63a7\u591a\u6a21\u614b\u6a21\u578b\u884c\u70ba"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-daea5ca9164c.jpg\" alt=\"MMDiff\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">MMDiff\uff08Multimodal Model Diffing\uff09\u4fc2\u4e00\u5957\u7814\u7a76\u578b\u65b9\u6cd5\uff0c\u8655\u7406\u5605\u554f\u984c\u5514\u4fc2\u518d\u8a13\u7df4\u6a21\u578b\uff0c\u800c\u4fc2\u5148\u7747\u6e05\u695a\u591a\u6a21\u614b\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08Multimodal Large Language Models, MLLMs\uff09\u5165\u9762\u908a\u5572\u7279\u5fb5\u56e0\u70ba\u8996\u89ba\u8a13\u7df4\u800c\u6539\u8b8a\uff0c\u518d\u5c07\u5462\u5572\u7279\u5fb5\u79fb\u9664\u6216\u5f15\u5c0e\u3002\u5c0d\u8981\u505a\u5be9\u6838\u3001\u9664\u932f\u3001\u5c0d\u9f4a\u884c\u70ba\u5605\u5718\u968a\uff0c\u5462\u7a2e\u505a\u6cd5\u6bd4\u55ae\u9760\u9ed1\u76d2\u8f38\u51fa\u66f4\u5be6\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f62\u5605\u6838\u5fc3\u505a\u6cd5\u4fc2\u5c07\u57fa\u790e\u8a9e\u8a00\u6a21\u578b\u5605 sparse autoencoder\uff08SAE\uff09\u540c\u591a\u6a21\u614b SAE \u5c0d\u9f4a\u6bd4\u8f03\uff0c\u518d\u7528 contrastive per-token firing \u53bb\u63c0\u51fa\u4efb\u52d9\u76f8\u95dc\u7279\u5fb5\u3002\u4e4b\u5f8c\u53ef\u4ee5\u505a causal removal\uff0c\u6216\u8005\u7528 MMDiff-CAA \u55ba\u76f8\u95dc\u5c64\u9762 steering\uff0c\u76f4\u63a5\u6539\u8b8a\u6a21\u578b\u55ba\u7a7a\u9593\u63a8\u7406\u3001OCR \u540c\u5b89\u5168\u5834\u666f\u5605\u8f38\u51fa\u50be\u5411\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u63d0\u4f9b\u5497\u5c0d\u61c9\u5605 SAE checkpoints\uff0c\u4ea6\u6709\u91dd\u5c0d LLaVA-MORE\u3001PaliGemma 2 \u540c InternVL3.5 \u5605\u7d50\u679c\u3002\u4f5c\u8005\u505a\u5497\u8996\u89ba\u7a7a\u9593\u7406\u89e3\u3001\u591a\u6a21\u614b\u5b89\u5168\u540c OCR \u8a55\u6e2c\uff0c\u6307\u51fa\u79fb\u9664\u55ae\u4e00\u7279\u5fb5\u53ef\u4ee4\u76ee\u6a19\u80fd\u529b\u4e0b\u964d\uff0c\u540c\u6642\u4e00\u822c VQA \u53d7\u5f71\u97ff\u597d\u7d30\uff0c\u986f\u793a\u5462\u5957\u65b9\u6cd5\u5514\u4fc2\u7d14\u89c0\u5bdf\uff0c\u800c\u4fc2\u771f\u4fc2\u53ef\u4ee5\u505a\u5b9a\u5411\u63a7\u5236\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53ef\u7528\u569f\u5b9a\u4f4d\u591a\u6a21\u614b\u8a13\u7df4\u6539\u5beb\u5497\u908a\u5572\u7279\u5fb5<\/li>\n<li>\u53ef\u4ee5\u5c07\u7279\u5b9a\u884c\u70ba\u524a\u5f31\uff0c\u800c\u5514\u4e00\u5b9a\u8981\u91cd\u8a13\u6574\u500b\u6a21\u578b<\/li>\n<li>\u5c0d\u505a OCR\u3001\u7a7a\u9593\u63a8\u7406\u3001\u6a21\u578b\u5b89\u5168\u5be9\u67e5\u5605\u5718\u968a\u7279\u5225\u6709\u7528<\/li>\n<li>\u6e2c\u8a66\u7d50\u679c\u986f\u793a\uff0c\u76ee\u6a19\u80fd\u529b\u4e0b\u964d\u660e\u986f\uff0c\u4e00\u822c VQA \u5e72\u64fe\u76f8\u5c0d\u7d30<\/li>\n<li>\u65b9\u6cd5\u53ef\u8de8\u5e7e\u500b MLLM \u5bb6\u65cf\u8f49\u7528\uff0c\u5514\u53ea\u9650\u55ae\u4e00\u6a21\u578b<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/hunarbatra.com\/MMDiff\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/hunarbatra\/MMDiff\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/OX-PIXL\/MMDiff_SAEs\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>MMDiff \u628a\u6587\u5b57\u6a21\u578b\u540c\u591a\u6a21\u614b\u7248\u672c\u4e4b\u9593\u5605\u5dee\u7570\u62c6\u958b\uff0c\u627e\u51fa\u53d7\u8996\u89ba\u8a13\u7df4\u6539\u5beb\u5605\u7279\u5fb5\uff0c\u518d\u7528\u569f\u524a\u5f31\u6216\u5f15\u5c0e\u6a21\u578b\u884c\u70ba\u3002\u5c0d\u505a\u591a\u6a21\u614b\u5b89\u5168\u3001\u7a7a\u9593\u63a8\u7406\u540c OCR \u5605\u5718\u968a\uff0c\u4f62\u63d0\u4f9b\u5497\u4e00\u689d\u53ef\u89e3\u91cb\u53c8\u53ef\u5e72\u9810\u5605\u8def\u3002<\/p>\n","protected":false},"author":8,"featured_media":11132,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,173,119],"tags":[],"class_list":["post-11133","post","type-post","status-publish","format-standard","hentry","category-133","category-173","category-119"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11133","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=11133"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11133\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11132"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11133"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11133"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11133"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}