Tsinghua's CoACT cuts Coding Agent tokens by 33% while slightly improving pass rate
机器之心 · wechat · 2026-11-13
A Tsinghua AI Institute team proposes CoACT, an observation-compression method for Coding Agents judged not by textual similarity but by whether the agent's next action changes (Next-Action Preservation). Accepted at ICSE 2027 (top 5%).
- Motivation: observations account for 45.7% of tokens on SWE-bench Verified and up to 67.8% on Terminal-Bench; CoACT compresses before insertion so KV cache stays reusable
- Training: Gemini3-Flash generates 8 compression candidates per observation; each is scored against 8 sampled reference actions (compared by operation type, target file, read range), keeping the shortest passing candidates; a Qwen3.5-4B compressor is trained in two stages (offline SFT + online alignment)
- Results: 33.0% average total-token reduction across three agent models, with average PASS@1 rising from 69.0% to 71.0%; interaction steps drop from 56.5 to 51.2
- Baselines like LLMLingua-2 inflate total tokens and hurt accuracy; CoACT is the cheapest method on both benchmarks
- Combined with AgentDiet, which alone raises estimated cost from $41.26 to $45.65 due to KV-cache invalidation, adding CoACT cuts cost to $25.88 (43.3% cheaper)
More from coding & agent
- When playing and making games become the same activity: agents rewrite game dev economics — TAbrodi · 2026-11-13
- Firecrawl CTO on growing a 1,000-star repo to 1.2M developers — garrytan · 2026-11-13
- Swapping one month of Scaleway bills for a 16GB ThinkCentre to run agents locally — jiayuan_jy · 2026-11-13
- StackShift launches cloud Linux machines with root, Docker and Chromium for AI agents — Scobleizer · 2026-11-13
- Agent-Native Clips enters open-source beta, giving agents eyes and ears via shared links — Scobleizer · 2026-11-13
- OHAN opens early access: a network where AI agents talk directly to agents — Scobleizer · 2026-11-13