ACE-RL: RL-trained agents learn to curate experience playbooks that boost future rollouts
kaist-ai · hf · 2026-11-14
A NYU/KAIST AI paper, "Experience-Handoff Reinforcement Learning for Recursive Self-Improvement," notes that scalar-reward RL discards rich environment information from trials (e.g., error logs), while reflection/compaction harnesses hand off experience without ever learning how to build the artifacts.
Their method ACE-RL fixes an Agentic Context Engineering (ACE) harness and trains role-specific behaviors — action generator, insight reflector, playbook curator — with RL, evaluating candidate experience-handoff artifacts by their downstream effect on future actor rollouts. This combines scalar-reward optimization with environment information, letting agents improve recursively by creating and curating experience artifacts.
More from coding & agent
- Your company 'second brain' is just silos with an AI on top — alex_verem · 2026-11-14
- Modal Endpoints land in Nous Research's Hermes Agent via one-line plugin install — AAAzzam · 2026-11-14
- Cmpnd releases dspy-wsp: $5-10 of GPT-6 edits model weights, lifts every decision model tested — dbreunig · 2026-11-14
- AmphoraDB launches serverless databases for agent swarms, paid per statement in stablecoins — banteg · 2026-11-14
- Ex-quant's local AI agent negotiated 1,000+ influencer outreach in 9 days — Shruti_0810 · 2026-11-14
- With 90% AI-generated code, Limestone gates merges on its V.U.E. three checks — alex_verem · 2026-11-14