💻 Self-Improving Coding Agents

We study how coding agents can improve their own decisions—and the strategies they use to improve. Our research focuses on curating high-quality (accurate and concise) trajectories for training coding agents: solving repository-level software engineering tasks with fewer unnecessary actions and better use of computation.

💡 Featured project: (Self-Improving)² Coding Agents

SI2CA explores self-improvement at two levels:

• ⚖️ Improve individual decisions. A model generates candidate actions and selects among them using Self-Judgement, which evaluates candidates against a rubric, or Self-Likelihood, which scores candidates using likelihood conditioned on privileged information.
• 🔄 Improve the improvement strategy. A recursive self-improvement (RSI) framework searches for better ways to decide when and how to branch, reducing the cost of curating high-quality trajectories.

🧬 Better data, better coding agents

We evaluate these methods on SWE-bench Verified, SWE-bench Pro, and DeepSWE, studying task success, trajectory conciseness, computational cost, and downstream training.

We also validate the effectiveness of our data curation through supervised fine-tuning of a smaller model, Qwen3.5-35B-A3B-Base. Compared with training on standard rollouts, training on Self-Judgement-curated trajectories yields higher task accuracy and fewer average turns on SWE-bench Verified and the Python subset of SWE-bench Pro. These results show that the curated trajectories provide better training data, helping a smaller coding agent solve more tasks with fewer steps.

🚀 Explore the project

• SI2CA — implementation, trajectory curation, evaluation, and training.