BSR-32B: Behavioral Self-Rewarding for Software Engineering Agents

BSR-32B is a software engineering agent based on Qwen3-32B, trained with Behavioral Self-Rewarding (BSR), a reinforcement learning method that improves upon binary verifiable rewards.

BSR enables the policy to evaluate its own debugging trajectories through pairwise behavioral comparisons, providing additional training signals even when trajectories share the same verification outcome. It improves software engineering performance without requiring a separate reward model or additional environment interactions.

Resources: 📄 Paper | 🌐 Project Page | 💻 GitHub | 🤗 BSR-14B

Training

• Base model: Qwen3-32B
• Training data: Filtered R2E-Gym dataset (4,578 tasks)
• Training method: Agentic reinforcement learning with LOOP and Behavioral Self-Rewarding
• Reward design: Binary verification rewards augmented with bounded, policy-generated behavioral rewards for failed trajectories

Evaluation

BSR consistently improves over verifier-only reinforcement learning on SWE-bench Verified and SWE-bench-Live Lite.

Benchmark	Baseline	BSR-32B
SWE-bench Verified (Avg. Acc.)	34.4%	35.5%
SWE-bench Verified (Pass@3)	48.8%	50.6%
SWE-bench-Live Lite (Avg. Acc.)	7.67%	8.56%
SWE-bench-Live Lite (Pass@3)	12.33%	13.67%

Results correspond to the 270-step training setting. Evaluation uses three runs per checkpoint, temperature 1.0, and a maximum of 50 agent steps. Additional results on Terminal-Bench 2.1 are available in the paper.

Usage

The model is intended for repository-level software engineering tasks with an interactive agent environment and developer tools. See the GitHub repository for the agent setup and evaluation framework.

🎈 Citation

If you use BSR, please cite our paper.

@article{wu2026beyond,
  title = {Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents},
  author = {Wu, Yiran and Zhao, Andrew and Chang, Yurui and Wu, Qingyun and Feng, Xuan},
  journal = {arXiv preprint arXiv:2610.12598},
  year = {2026}
}