SWE-Trail uses history to guide coding agent trajectory collection
Original title: SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training
SWE-Trail improves SWE-bench resolution by 2.33 points over independent sampling for a 124B student model.
Coding agents collect trajectories through exploration, editing and validation. These trajectories shape post-training data for models. SWE-Trail uses earlier trajectories to identify states where additional collection effort may help. The coding agent then proposes candidate actions from its ordinary context. This approach avoids train-test mismatch from exposing history to the agent. For a 124B student initialized from Ling-3.0-flash-base, SWE-Trail improves mean resolution by 2.33 points on SWE-bench Multilingual and 1.60 points on SWE-bench Pro.
- SWE-Trail identifies states needing more collection effort
- Agent proposes actions from ordinary context only
- Improves 124B model on SWE-bench Multilingual by 2.33 points
- Improves on SWE-bench Pro by 1.60 points
From the original article
1 min read at huggingface.co
Coding agents solve repository issues through long sequences of exploration, editing, and validation. These trajectories are increasingly used as post-training data, so how they are collected shapes what the trained model learns. A common approach is to sample multiple trajectories independently for each issue and then select or curate them.
By Ziming Li, Lu Yu, Jiatan Huang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.