SWE-Trail Guides Coding-Agent Training Data Collection
SWE-Trail is a trajectory collector that uses earlier coding-agent trajectories to decide where additional collection effort is useful, while keeping that history outside the agent’s context. In the authors’ experiments, a 124B student initialized from Ling-3.0-flash-base improved mean resolution over independent sampling by 2.33 percentage points on SWE-bench Multilingual and 1.60 points on SWE-bench Pro.
The method offers a way to reduce redundant training trajectories without creating a train-test mismatch from adding unavailable history to an agent’s prompt.
Computer Science > Artificial Intelligence
arXiv:2611.02074 (cs)
[Submitted on 1 Nov 2026]
Title:SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training
Authors:Ziming Li, Lu Yu, Jiatan Huang, Qing Cui, Jun Zhou, Chuxu Zhang
View a PDF of the paper titled SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training, by Ziming Li and 5 other authors
#SWE-Trail#Ling-3.0-flash-base#SWE-bench Multilingual#SWE-bench Pro