[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-6b52b4a9-d567-46b8-99c1-e9c65ba59b16":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},"6b52b4a9-d567-46b8-99c1-e9c65ba59b16","SWE-Pruner Pro:ByteDance 让 Agent 自己当剪枝器,省 39% token 还涨分","多轮编程 Agent 跑长任务时,上下文里 80% 是工具返回的「无用行」——测试日志、报错堆栈、diff 残留。SWE-Pruner Pro 反直觉:哪些行该删,Agent 自己的中间层表征已经知道,根本不需要外挂分类器。\\n\\n字节 Seed 这篇 arXiv(2607.18213)在冻结 Coder LLM 上挂小剪枝头,把 Agent 阅读工具输出时的 hidden state 映射成「行级 keep\u002Fprune」决策,叠 length-aware embedding 区分不同长度工具块。在 MiMo-V2-Flash 等两个开源 backbone、四个多轮基准上几乎不增加推理延迟就省下最多 39% 的 prompt+completion token,SWE-Bench Verified 求解率 +3.8pp,Oolong 长上下文准确率 +2.2pp。\\n\\n意义在范式转移。之前所有上下文压缩方法都默认「需要独立判定模型」,Pro 用 Coder 自身表征证明这是多余假设——Agent 已是「上下文相关性」最优判断器,只是没人把这层信号读出来。代码已开源,下一步:能否搬到通用 Agent。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.18213","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",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},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-07-25T12:00:00Z","2026-07-25T12:06:52.414195Z","2026-07-25T12:06:52.414206Z",true,"agent",3]