SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training
Story 9, from arXiv
SWE-Trail is a new method for collecting coding agent trajectories that uses past data to guide future collection, improving performance without exposing history to the model during training.
Key points
- SWE-Trail uses earlier trajectories to guide where new collection effort should be focused.
- History is kept outside the coding agent's context to avoid train-test mismatch.
- The method identifies useful states for additional exploration during a new rollout.
- SWE-Trail improves resolution on SWE-bench benchmarks compared to independent sampling.
- Without history, the same pipeline performs worse than independent sampling on some benchmarks.
Why it matters
This approach could help improve the efficiency and effectiveness of training coding agents by making better use of historical data, which is critical for tasks like automated bug fixing and code generation.
Counterpoint
The method relies on having a sufficient amount of historical data, which may not be available for all coding tasks or repositories.
Sources
This summary was written by a language model and may contain errors. Read the original for the full story.