Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Tsinghua's CoACT cuts coding-agent token use by 33%

AI summary

Machine Intelligence reports that researchers from Tsinghua University proposed CoACT, an observation-compression method that preserves a coding agent’s next action rather than measuring similarity to the original observation. On 200 SWE-bench Verified instances with three agent models, CoACT reportedly reduced total token consumption by 33.0% on average, while average PASS@1 rose from 69.0% to 71.0% and interaction steps fell from 56.5 to 51.2. The study reports that CoACT can also be combined with AgentDiet, with estimated total cost falling to $25.88 while task success remained similar.

In-depth insight

CoACT offers a concrete way to reduce the context overhead of coding agents while preserving downstream behavior, and the reported results cover multiple agent models rather than a single setup.

机器之心 2026-11-13 14:28 北京

能否在不影响任务完成的前提下压缩上下文?

Coding Agent 的上下文越长,每一步调用的开销越高,能否在不影响任务完成的前提下压缩上下文?

Coding Agent 解决软件问题时,需要反复查看代码、修改程序、运行测试:每一步返回的文件内容、搜索结果和日志,都会留在上下文里,下一轮调用时还要整段重新带上。任务越难,上下文越长,每一步要处理的 token 就越多,费用也越高。统计发现,在 SWE-bench Verified 上,这些 observation 占了总 token 消耗的 45.7%,在 Terminal-Bench 上最高达到 67.8%。

#CoACT#Tsinghua University#Qwen3.5-4B#Qwen3.5-35B-A3B#Deepseek-v4-Pro#Gemini3-Flash#AgentDiet#LLMLingua-2#SWE-Pruner#LongCodeZip#SWE-bench Verified

Original · 6 min readRead the original
Group chat

No comments yet. Start the conversation.