RECAP: RL recipe for context compaction in long-horizon agents, +4.3pp on SWE-bench
Michael_D_Moor · x · 2026-11-05
An ETH/EPFL collaboration led by Yulun Jiang introduces RECAP, an RL recipe for training long-horizon LLM agents with context compaction.
- Core idea: A good summary should preserve the agent's behavior under the original interaction history. RECAP combines final task rewards with a measure of how compaction changes agent behavior.
- Results: BrowseComp-Plus improves from 58.2% to 62.8% (+4.6pp) and SWE-bench Verified from 46.8% to 51.1% (+4.3pp), with fewer tool calls and shorter trajectories.
- Paper and code are publicly available.
More from coding & agent
- Dev finds pi coding tool defaults to Opus 4.8 on a fresh install — lucasmeijer · 2026-11-05
- Kylon puts whole teams and agents in shared rooms, with a Neon branch per app draft — jefrankle · 2026-11-05
- "Evals > vibes": agent workflow debates are just guessing without eval stats — annetgriffin · 2026-11-05
- OpenAI hackathon winner joins YC, launches Raven, an agent-first phone — ycombinator · 2026-11-05
- Open-source Cicada brings real-time translation and AI meeting notes to any Mac call, free 50h/month — dotey · 2026-11-05
- IBM launches AI coding assistant named "Bob", devs can't stop laughing — zeeg · 2026-11-05