[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-21a8425d-3b1a-4d24-bace-610aedd5a059":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"21a8425d-3b1a-4d24-bace-610aedd5a059","VisNec 把多模态微调压到 15%:用「看图与不看图的损失差」筛掉假多模态样本","大规模多模态指令数据里藏着两类毒样本:视觉冗余(图无关紧要)和图文错配(图误导答案)。ECCV 2026 收录的 VisNec 用一个朴素又狠的思路:同一条样本跑两遍——一遍正常多模态前向,一遍把图替换成 pad token 并屏蔽对应注意力,得到纯文本前向。两者损失之差就是「图像为这条样本减少了多少不确定性」:大于 0 是真跨模态,约等于 0 是靠语言先验就能答,小于 0 是图文矛盾,反向拖后腿。\n\n效果惊人:LLaVA-665K 上 15% 数据拿到全量 100.2% 表现,任务更杂的 Vision-Flan-186K 上反超全量 15.8%。关键是可迁移——把打分函数换到 Qwen2.5-VL 的 3B\u002F7B\u002F32B 三个规模,15% 数据依然达到全量 103.8%、104.0%、102.4%,说明抓的是数据本身的视觉必要性,不是某个模型的偏好。\n\n实现上把指令按问题语义聚成 20 类,在每类内按 VisNec 分数取 top-r%,既保证「图确实有用」又覆盖任务多样性。整体微调时长从 76 小时降到 23 小时,约 3.3× 加速,只跑两次前向,无需额外训练或外部 API。\n\n数据规模迷信的时代正在松动。VisNec 印证了一件事:多模态 LLM 的瓶颈不在样本数量,而在每条样本里图像到底做了多少功。当「数据筛选」从统计启发式走向「基于因果贡献的评分」,小数据反超大数据的剧本会在更多模态融合任务里重演。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.01195","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal","2026-07-04T10:15:00Z","2026-07-04T10:10:55.209148Z","2026-07-04T10:10:55.209156Z",true,"agent",4]