论文精选

UCLA 与 Northeastern 论文:AI 智能体约 18% 输出 token 属于可删废料

精选理由

跑过 agent 的都知道账单有多吓人。这篇论文用 LLM judge 找出哪些 token 是纯浪费,Kimi-K3 在 SWE-bench Verified 上删掉 41% 输出成功率还微升。做 agent 成本优化的开发者可以照这个思路搭个代理层试试。

UCLA 与 Northeastern University 的论文测算,AI 智能体的输出 token 中约 18% 是可移除的废料,删掉后任务成功率几乎不受影响。传统按总 token 计费的方式无法区分哪些 token 真正做了有用工作,哪些只是在重复策略、重新拉取文件或走入死胡同。

正文

研究方法是把智能体调用路由经过一个代理层,由 LLM judge 标注可删除的文本片段,然后删掉这些片段并重跑智能体验证任务是否正常。实验覆盖 6 个前沿模型和 4 个基准。对 Kimi-K3 的推理过程做重复与填充词修剪后,在 SWE-bench Verified 上删除了 41% 的输出,成功率从 89.0% 变为 89.5%,几乎持平。

对开发者而言,降低智能体成本的关键是定位任务和 harness 产生的浪费来源,而不是简单设置 token 上限。

原文 · rohanpaul_ai

New UCLA+ Northeastern University paper finds that about 18% of AI agent output tokens are removable waste, and cutting it barely hurts task success.

Agent bills are usually measured as total tokens. That number can't tell you which tokens did useful work and which just repeated a policy, refetched a file, or went down a dead end.

They route agent calls through a proxy and have an LLM judge mark spans that can go without losing useful work. Then they delete those spans and rerun the agent to check nothing breaks.

Across 6 frontier models and 4 benchmarks, about 18% of output tokens were waste. Pruning repetition and filler from Kimi-K3's reasoning removed 41% of it on SWE-bench Verified, and success stayed at 89.5% versus 89.0%.

To cut agent costs, find where your task and harness create waste instead of just capping tokens.

产业坐标与信源核验

本条情报归属 AITOP「论文」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」「中美大模型技术代差与商业化对照」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:Kimi-K3、SWE-bench Verified、智能体、LLM judge,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明:本文原始线索采编自「rohanpaul_ai」(@rohanpaul_ai),经 AITOP 多维评分管线结构化提炼。