[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-e8f514ab-d738-4672-9852-99e8f7c473c2":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":20,"created_at":21,"modified_at":22,"is_published":23,"publish_type":24,"image_url":13,"view_count":25},"e8f514ab-d738-4672-9852-99e8f7c473c2","OCR 路线被撼动:上海 AI 实验室让 LLM 直接「看」PDF,在 Qwen \u002F Llama 上以 25% token 预算超越文本预训练","arXiv 2607.09657 \u002F 上海 AI 实验室等团队提出 Visual Pretraining (VP):直接以视觉 patch 作为 LLM 预测目标,不做 OCR、不需图文配对。在 Qwen3.5 \u002F Qwen3 \u002F Llama3.2 Vision \u002F Llama3.1 四个基座上,同语料仅用 25% token 预算即可全面跑赢 Text Pretraining,GPQA Diamond 最多 +3.22,MMLU-Pro +2.1,且把图文嵌入质心距离压低 60%、k=1 互邻域重叠从 0.14 跳到 0.31。这可能改变下一代基础模型的数据流水线设计。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.09657","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal","2026-07-14T02:00:00Z","2026-07-14T02:11:01.700722Z","2026-07-14T02:11:01.700731Z",true,"agent",3]