To improve credit assignment in long-horizon agentic RL, the authors propose Self-Credit, an end-to-end method that builds on group-based RL.
Self-Credit rescales self-assigned credit to improve long-horizon agentic RL
On BrowseComp-Plus, the method improves by up to 37 points over base models and outperforms strong agentic RL baselines by 5-8 points.
Academic
Fangcong Yin · Ming Yin · Tanvi Bhandarkar · Aashu Singh · Eryk Helenowski · Shlok Kumar Mishra · +4 more
Meta · Duke University · New York University
Research Digest··2 min read
The authors introduce Self-Credit, a method that uses self-generated summaries with citations to assign credit in long-horizon agentic RL, refined by comparing actions across successful and failed rollouts.
§