31 upvotes
History-Guided Trajectories Boost Training of Code-Fixing Agents
What is SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training?
SWE-Trail proposes a way to collect the step-by-step actions that a coding assistant takes when fixing bugs, using the repository’s past commits to steer the process. The resulting trajectories are richer and more realistic than independently sampled ones, helping models learn better editing and validation behavior.
Who it is for
Researchers and teams building AI code assistants should look at it to improve their training data and potentially get higher success rates on real issues.
How to try it
Open the paper on Hugging Face: https://huggingface.co/papers/2611.02074