{"id":10095,"date":"2026-07-10T12:46:55","date_gmt":"2026-07-10T04:46:55","guid":{"rendered":"https:\/\/infernews.com\/blog\/splash\/"},"modified":"2026-07-10T12:48:16","modified_gmt":"2026-07-10T04:48:16","slug":"splash","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/splash\/","title":{"rendered":"SPLASH \u8b93\u5c0f\u578b MLLM \u5b78\u6703\u89f8\u89ba"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/figure.jpg\" alt=\"figure\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u60f3\u5c07\u89f8\u89ba\u8cc7\u6599\u63a5\u5165\u5c0f\u578b\u591a\u6a21\u614b\u6a21\u578b\uff0c\u6700\u9ebb\u7169\u7684\u5730\u65b9\u5f80\u5f80\u4e0d\u662f\u63a5\u5514\u63a5\u5230\uff0c\u800c\u662f\u6a21\u578b\u4e00\u908a\u5b78\u300c\u6478\u5230\u4e5c\u300d\uff0c\u4e00\u908a\u628a\u539f\u672c\u300c\u7747\u5230\u4e5c\u300d\u7684\u80fd\u529b\u641e\u4e82\u3002SPLASH\u5c6c\u65bc<strong>\u6a21\u578b\u8a13\u7df4\u6846\u67b6<\/strong>\uff0c\u91dd\u5c0d\u7684\u6b63\u662f MLLMs \u5728\u52a0\u5165 tactile perception \u5f8c\u5bb9\u6613\u51fa\u73fe\u7684 catastrophic forgetting\uff0c\u76ee\u6a19\u662f\u5728\u4fdd\u4f4f vision-language \u80fd\u529b\u4e4b\u4e0b\u5b8c\u6210 visuo-tactile \u5c0d\u9f4a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u53d6\u5411\u5e7e\u6e05\u695a\uff1a\u5514\u4fc2\u5168\u9762\u91cd\u8a13\uff0c\u4e5f\u5514\u4fc2\u96a8\u4fbf\u52a0\u4e00\u689d tactile branch \u5c31\u7b97\uff0c\u800c\u662f\u5148\u5728 LLM backbone \u5165\u9762\u627e\u51fa\u8f03\u300c\u6c89\u7761\u300d\u7684\u53c3\u6578\u7a7a\u9593\uff0c\u518d\u628a\u89f8\u89ba\u5b78\u7fd2\u9650\u5236\u5728\u5462\u90e8\u5206\u3002\u9805\u76ee\u63d0\u5230\u5b83\u7528 weight \u8207 activation importance scoring \u751f\u6210 dormant masks\uff0c\u4e4b\u5f8c\u505a mask-guided sparse training\uff1b\u597d\u8655\u662f\u5514\u4f7f\u5927\u5e45\u52d5\u5230\u95dc\u9375\u8996\u89ba\u8a9e\u8a00\u53c3\u6578\uff0c\u4ee3\u50f9\u5247\u662f\u6574\u500b\u6d41\u7a0b\u4ecd\u7136\u504f\u7814\u7a76\u578b\uff0c\u90e8\u7f72\u524d\u8981\u5148\u5099\u597d\u8cc7\u6599\u96c6\u3001\u5206\u5272\u8cc7\u6599\uff0c\u540c\u6642\u9700\u8981 CUDA 12.0 \u4ee5\u4e0a\u8207\u81f3\u5c11\u5169\u5f35 GPU \u505a\u5206\u6563\u5f0f\u8a13\u7df4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73fe\u6709\u7248\u672c\u4e3b\u8981\u6709\u5169\u500b\u6a21\u578b\u8b8a\u9ad4\uff1aSPLASH-1B \u4ee5 InternVL2.5-1B \u70ba base MLLM\uff0cSPLASH-3B \u5247\u5efa\u57fa\u65bc Qwen2.5-VL-3B\uff0c\u5169\u8005\u90fd\u914d\u5408 ViT-Tiny + MLP adapter \u4f5c tactile frontend\u3002\u8cc7\u6599\u90e8\u5206\u4ea6\u5514\u7b97\u8f15\u91cf\uff0c\u9664\u4e86 LLaVA-CC3M-Pretrain-595K \u8207 CC3M \u7528\u4f86\u751f\u6210 mask\uff0c\u4ef2\u8981\u914d\u5408 Touch-Vision-Language-Dataset\u3001TacQuad \u7b49\u9805\u76ee\u505a\u8a13\u7df4\u8207 OOD \u8a55\u4f30\uff0c\u6240\u4ee5\u5b83\u6bd4\u8f03\u9069\u5408\u505a\u591a\u6a21\u614b\u7814\u7a76\u3001\u6a5f\u68b0\u611f\u77e5\u3001\u6216\u60f3\u9a57\u8b49\u89f8\u89ba\u2014\u8996\u89ba\u806f\u5408\u63a8\u7406\u7684\u5718\u968a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img data-dominant-color=\"dadbda\" data-has-transparency=\"false\" style=\"--dominant-color: #dadbda;\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"284\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/image-1024x284.png\" alt=\"\" class=\"wp-image-10097 not-transparent\" srcset=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/image-1024x284.png 1024w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/image-300x83.png 300w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/image-768x213.png 768w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/image.png 1100w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u91cd\u9ede\u4e0d\u5728\u65b0\u589e\u591a\u5c11\u53c3\u6578\uff0c\u800c\u5728<strong>\u628a\u89f8\u89ba\u66f4\u65b0\u9694\u96e2\u5230 dormant subspace<\/strong><\/li>\n\n\n\n<li>\u57fa\u5e95\u6a21\u578b\u5305\u62ec <strong>InternVL2.5-1B<\/strong> \u8207 <strong>Qwen2.5-VL-3B<\/strong><\/li>\n\n\n\n<li>\u89f8\u89ba\u524d\u7aef\u63a1\u7528 <strong>ViT-Tiny + MLP adapter<\/strong><\/li>\n\n\n\n<li>\u8a55\u4f30\u8986\u84cb <strong>SSVTP\u3001TVL\u3001TacQuad<\/strong>\uff0c\u4e26\u5f37\u8abf\u4fdd\u7559\u539f\u6709\u901a\u7528\u80fd\u529b<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5\u516c\u958b\u8cc7\u8a0a\u5224\u65b7\uff0cSPLASH\u6700\u503c\u5f97\u7559\u610f\u7684\u4e0d\u662f\u55ae\u4e00 benchmark \u5206\u6578\uff0c\u800c\u662f\u5b83\u660e\u78ba\u62bc\u6ce8\u300cnon-destructive modality expansion\u300d\uff1a\u8b93\u6a21\u578b\u591a\u5b78\u4e00\u7a2e\u611f\u5b98\uff0c\u800c\u5514\u9700\u8981\u7528\u8996\u89ba\u80fd\u529b\u505a\u4ea4\u63db\u3002\u5c0d\u6253\u7b97\u5728\u7dca\u6e4a\u53c3\u6578\u9810\u7b97\u4e0b\u64f4\u5c55 MLLMs \u611f\u77e5\u6a21\u614b\u7684\u7814\u7a76\u9805\u76ee\u800c\u8a00\uff0c\u9019\u500b\u65b9\u5411\u6bd4\u55ae\u7d14\u8ffd\u9ad8\u8868\u73fe\u66f4\u6709\u53c3\u8003\u50f9\u503c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"http:\/\/mmai.ewha.ac.kr\/splash\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/ewha-mmai\/splash\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>SPLASH\u60f3\u8655\u7406\u7684\uff0c\u4e0d\u662f\u55ae\u7d14\u52a0\u4e00\u500b\u611f\u6e2c\u6a21\u7d44\uff0c\u800c\u662f\u4ee4\u5c0f\u578b\u591a\u6a21\u614b\u6a21\u578b\u5b78\u89f8\u89ba\u6642\u5514\u6703\u5fd8\u8a18\u539f\u672c\u7684\u8996\u89ba\u8a9e\u8a00\u80fd\u529b\u3002\u4f60\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\u4e00\u7a2e\u6709\u7bc0\u5236\u7684\u5fae\u8abf\u65b9\u6cd5\uff0c\u5c08\u9580\u907f\u958b\u707d\u96e3\u6027\u907a\u5fd8\u3002<\/p>\n","protected":false},"author":8,"featured_media":10094,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,179,119,127,199],"tags":[],"class_list":["post-10095","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-nvidia","category-119","category-127","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10095","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=10095"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10095\/revisions"}],"predecessor-version":[{"id":10098,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10095\/revisions\/10098"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10094"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10095"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10095"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10095"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}