Hugging Face Releases FC-SWE: Revolutionizing Software Engineering Agent Training with Failure-Conditioned RL

Summary:Hugging Face introduces FC-SWE, a novel failure-conditioned reinforcement learning (RL) framework designed to train long-horizon software engineering agents. By leveraging failed patches and verifier feedback as context for recovery trajectories, FC-SWE significantly enhances agent performance in complex tasks. In the SWE-bench benchmark, FC-SWE, powered by Qwen3.5-4B and SWE-agent, achieved a 41.7% Resolved@1 and 52.8% Resolved@2, outperforming the traditional GRPO method by 2.8% and 4.3%, resp

---

Technical Mechanism Analysis

FC-SWE (Failure-Conditioned Software Engineering) is a reinforcement learning (RL) framework designed to address the limitations of traditional RL methods in handling long-horizon software engineering tasks. Its core mechanisms include:

1. Failure-Conditioned Recovery: After a patch fails verification, FC-SWE restores the repository to its original task state and uses the failed patch and reviewer feedback as context to generate a recovery trajectory. This ensures the agent learns from failures rather than repeating mistakes.

2. Trajectory-Local Rewards: FC-SWE introduces a trajectory-local reward mechanism to ensure that the reviewer outcome of each attempt does not affect the reward of earlier failed patches. This prevents recovery success from rewarding an earlier failed patch, avoiding confusion during training.

3. Active-Set Advantage Estimation: By comparing all actually executed initial and recovery trajectories for the same issue, FC-SWE forms an active-set advantage estimation group, ensuring failed attempts remain in the group while unexecuted attempts are excluded. This improves training efficiency and reduces the impact of ineffective attempts on the model.

Engineering Trade-offs and Performance

FC-SWE was tested on 500 tasks in SWE-bench, demonstrating a Resolved@1 and Resolved@2 rate of 41.7% and 52.8%, respectively, when combined with Qwen3.5-4B and SWE-agent. This is a significant improvement over the traditional GRPO method, which achieved 38.9% and 48.5%. Despite being trained with at most two attempts per chain, FC-SWE reached a Resolved@11 of 70.7% under an eleven-attempt test-time budget.

This performance gain comes at the cost of increased reliance on reviewer feedback and the potential complexity introduced in generating recovery trajectories. Additionally, the success of FC-SWE is contingent on the quality and timeliness of the reviewer feedback, which may pose challenges in certain application scenarios.

Developer Implementation and Deployment Recommendations

For developers looking to adopt FC-SWE, the following recommendations may be helpful:

• Reviewer Feedback Quality: Ensure the quality and timeliness of reviewer feedback to enhance the effectiveness of recovery trajectories.

• Resource Management: Given that FC-SWE requires additional computational resources to handle recovery trajectories, it is advisable to use it cautiously in resource-constrained environments.

• Model Selection: FC-SWE can be combined with various RL models; it is recommended to choose the appropriate model architecture based on the specific task.

• Continuous Optimization: Continuously monitor and optimize the reviewer feedback mechanism to further enhance FC-SWE's performance.

Conclusion

FC-SWE significantly improves the performance of software engineering agents in complex tasks by introducing a failure-conditioned recovery mechanism. Its impressive performance in the SWE-bench benchmark demonstrates the framework's great potential in AI-assisted software development. However, developers need to weigh the computational complexity and reliance on reviewer feedback based on their specific application scenarios.

---

Source: Hugging Face Daily Papers (2026-10-06)

— END —

Tags: #Hugging Face Follow #Reinforcement Learning Follow #Software Engineering Agents Follow #RL Framework Follow