FC-SWE trains coding agents on recovery after failed patches
Original title: FC-SWE: Failure-Conditioned RL for Long-Horizon Software Engineering Agents
With Qwen3.5-4B and SWE-agent it reports 41.7% Resolved@1 versus 38.9% for GRPO.
FC-SWE is a failure-conditioned reinforcement learning method for long-horizon software engineering agents. After verification fails, it restores the repository and conditions a recovery trajectory on the failed patch and verifier feedback. On all 500 SWE-bench Verified tasks under a verifier-assisted protocol, the authors report higher resolve rates than GRPO.
Unconfirmed: Scores are author-reported under a verifier-assisted protocol; independent replication is not stated.
- Resolved@2 is 52.8% for FC-SWE versus 48.5% for GRPO.
- Training uses at most two attempts per chain.
- At an eleven-attempt test budget, Resolved@11 is 70.7%.
- Trajectory-local rewards keep recovery success from crediting a failed patch.
- Products
- SWE-agentSWE-bench Verified
- AI models
- Qwen3.5-4B
From the original article
1 min read at huggingface.co
Repository-level software engineering (SWE) is a challenging long-horizon setting: agents must reason over extended interactions, use tools, and adapt to stateful environments.
By Jia Liufu, Bin Hu, Linglin Jing · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.