
{"id":8633,"date":"2026-05-30T18:55:13","date_gmt":"2026-05-30T10:55:13","guid":{"rendered":"https:\/\/infernews.com\/blog\/cvpr-2026-earlytom-early-token-compression-completes-fast-video-understanding\/"},"modified":"2026-05-30T18:55:13","modified_gmt":"2026-05-30T10:55:13","slug":"cvpr-2026-earlytom-early-token-compression-completes-fast-video-understanding","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/cvpr-2026-earlytom-early-token-compression-completes-fast-video-understanding\/","title":{"rendered":"EarlyTom \u4ee4\u5f71\u7247\u7406\u89e3\u6a21\u578b\u8dd1\u5f97\u66f4\u5feb"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/pasted-cc6d129c7727.jpg\" alt=\"earlytom logo\"><\/figure>\n<p>EarlyTom \u662f\u4e00\u500b\u91dd\u5c0d Video Large Language Models\uff08Video-LLMs\uff09\u800c\u8a2d\u7684 token \u58d3\u7e2e\u9805\u76ee\uff0c\u76ee\u6a19\u5f88\u6e05\u6670\uff1a\u5728\u4e0d\u91cd\u65b0\u8a13\u7df4\u6a21\u578b\u7684\u524d\u63d0\u4e0b\uff0c\u6e1b\u5c11\u5f71\u7247\u7406\u89e3\u6642\u8981\u8655\u7406\u7684\u8996\u89ba token \u6578\u91cf\u3002\u5b83\u4e3b\u8981\u8655\u7406\u63a8\u7406\u901f\u5ea6\u6162\u3001\u8a08\u7b97\u91cf\u9ad8\u9019\u500b\u75db\u9ede\uff0c\u5c24\u5176\u662f\u5f71\u7247\u5167\u5bb9\u6bd4\u5716\u7247\u66f4\u9577\u3001\u66f4\u91cd\uff0c\u6a21\u578b\u5f88\u5bb9\u6613\u5728\u524d\u6bb5\u7de8\u78bc\u5c31\u8017\u6389\u5927\u91cf\u6642\u9593\u3002<\/p>\n<p>\u9019\u500b\u9805\u76ee\u7684\u6838\u5fc3\u60f3\u6cd5\uff0c\u4e0d\u662f\u7b49\u8cc7\u6599\u5168\u90e8\u9032\u5165\u6a21\u578b\u5f8c\u624d\u58d3\u7e2e\uff0c\u800c\u662f\u66f4\u65e9\u5728 vision encoder \u968e\u6bb5\u52d5\u624b\u3002EarlyTom \u6703\u5229\u7528\u65e9\u671f transformer layers \u7684 attention \u8a0a\u865f\uff0c\u5148\u627e\u51fa\u5197\u9918 token\uff0c\u518d\u505a\u88c1\u526a\uff1b\u53e6\u6709\u53ef\u9078\u7684 inner compression\uff0c\u6703\u5728 LLM backbone \u6307\u5b9a\u5c64\u4ee5 DPC-KNN \u505a token \u5408\u4f75\u3002\u9019\u7a2e\u65e9\u671f\u58d3\u7e2e\u65b9\u5f0f\uff0c\u91cd\u9ede\u5728\u65bc\u9023 vision encoder \u7684\u8ca0\u64d4\u90fd\u4e00\u4f75\u6e1b\u5c11\u3002<\/p>\n<p>\u4f7f\u7528\u4e0a\uff0c\u9805\u76ee\u662f\u5efa\u57fa\u65bc LLaVA-NeXT\uff0c\u4e26\u53ef\u5305\u88dd LLaVA-OneVision \u6a21\u578b\uff1b\u7a0b\u5f0f\u5c64\u9762\u662f\u628a\u5df2\u8f09\u5165\u7684\u6a21\u578b\u518d\u5957\u7528 EarlyTom\u3002\u5c0d\u5df2\u7d93\u5728\u8dd1\u5f71\u7247\u554f\u7b54\u3001\u5f71\u7247\u63cf\u8ff0\u6216\u591a\u6a21\u614b\u7406\u89e3\u6d41\u7a0b\u7684\u4eba\u4f86\u8aaa\uff0c\u9019\u4ee3\u8868\u53ef\u5728\u539f\u6709\u6a21\u578b\u7ba1\u7dda\u4e0a\u52a0\u5165\u58d3\u7e2e\u6a5f\u5236\uff0c\u800c\u4e0d\u4e00\u5b9a\u8981\u6539\u52d5\u6574\u500b\u8a13\u7df4\u6d41\u7a0b\u3002<\/p>\n<ul>\n<li>\u514d\u8a13\u7df4\uff08training-free\uff09\u58d3\u7e2e\uff0c\u90e8\u7f72\u9580\u6abb\u76f8\u5c0d\u8f03\u4f4e<\/li>\n<li>\u5206\u70ba outer compression \u8207 optional \u7684 inner compression<\/li>\n<li>\u91cd\u9ede\u6539\u5584 Time-to-First-Token\uff08TTFT\uff09\u3001throughput \u8207 FLOPs<\/li>\n<li>\u4f9d\u8cf4\u65e9\u671f attention \u8a0a\u865f\u6311\u8d70\u5197\u9918\u8996\u89ba token<\/li>\n<li>\u76f8\u95dc\u6a21\u578b\u8207\u6846\u67b6\u5305\u62ec LLaVA-NeXT\u3001LLaVA-OneVision\u3001Qwen2<\/li>\n<\/ul>\n<p>\u5f9e\u516c\u958b\u8cc7\u6599\u770b\uff0c\u4f5c\u8005\u4ee5 lmms-eval \u9032\u884c\u8a55\u6e2c\uff0c\u4e26\u5728\u8ad6\u6587\u5167\u5bb9\u63d0\u5230\u6703\u6bd4\u8f03 MVBench\u3001EgoSchema\u3001LongVideoBench \u8207 VideoMME \u7b49\u5e38\u898b\u5f71\u7247\u7406\u89e3\u57fa\u6e96\u3002\u7d50\u679c\u63cf\u8ff0\u986f\u793a\uff0c\u5b83\u5728\u7dad\u6301\u63a5\u8fd1 full-token \u65b9\u6cd5\u6e96\u78ba\u5ea6\u7684\u540c\u6642\uff0cTTFT \u6700\u591a\u53ef\u964d\u81f3 2.65\u00d7\uff0c\u4ea6\u6709\u66f4\u9ad8 throughput\uff1b\u4e0d\u904e\u4e0d\u540c\u6a21\u578b\u5927\u5c0f\u3001\u5f71\u7247\u9577\u5ea6\u8207\u786c\u4ef6\u914d\u7f6e\u4e0b\uff0c\u5be6\u969b\u589e\u76ca\u4ecd\u8981\u5206\u958b\u770b\u3002<\/p>\n<p>\u9019\u500b\u9805\u76ee\u8f03\u9069\u5408\u5df2\u7d93\u4f7f\u7528 Video-LLMs \u7684\u7814\u7a76\u8005\u3001\u5de5\u7a0b\u5718\u968a\uff0c\u6216\u60f3\u5728\u8cc7\u6e90\u6709\u9650\u74b0\u5883\u4e2d\u63d0\u5347\u5f71\u7247\u7406\u89e3\u6548\u7387\u7684\u4eba\u3002\u82e5\u4f60\u95dc\u5fc3\u7684\u662f\u6a21\u578b\u7b54\u5f97\u6e96\u4e4b\u9918\uff0c\u4e5f\u8981\u66f4\u5feb\u958b\u59cb\u8f38\u51fa\u7d50\u679c\uff0cEarlyTom \u7684\u50f9\u503c\u5c31\u5728\u65bc\u5b83\u628a\u58d3\u7e2e\u6642\u6a5f\u63d0\u65e9\uff0c\u76f4\u63a5\u91dd\u5c0d\u6700\u82b1\u6642\u9593\u7684\u90e8\u5206\u4e0b\u624b\u3002<\/p>\n<p><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/viridisGreen\/EarlyTom\" rel=\"noopener noreferrer\">https:\/\/github.com\/viridisGreen\/EarlyTom<\/a><\/p>\n<p><strong>\u9805\u76ee\uff1a<\/strong> <a href=\"https:\/\/viridisgreen.github.io\/EarlyTom\/\" rel=\"noopener noreferrer\">https:\/\/viridisgreen.github.io\/EarlyTom\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>EarlyTom \u4ee5\u514d\u8a13\u7df4\u65b9\u5f0f\u6e1b\u5c11\u5f71\u7247\u63a8\u7406\u6642\u6240\u9700\u7684\u8996\u89ba token\uff0c\u91cd\u9ede\u662f\u52a0\u5feb\u63a8\u7406\u800c\u76e1\u91cf\u4fdd\u4f4f\u8868\u73fe\u3002\u5c0d Video-LLMs \u4f7f\u7528\u5834\u666f\u5c24\u5176\u6709\u53c3\u8003\u50f9\u503c\u3002<\/p>\n","protected":false},"author":8,"featured_media":8632,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,171,149,197],"tags":[],"class_list":["post-8633","post","type-post","status-publish","format-standard","hentry","category-133","category-171","category-149","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8633","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=8633"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8633\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8632"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8633"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8633"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8633"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}