SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training
| Source: arXiv AI
Tags: SWE-Trail, SWE-bench Pro, SWE-bench Multilingual, Ling-3.0, Coding agents, Trajectory collection
SWE-Trail uses earlier coding-agent trajectories to decide where to branch new rollouts without showing that history to the agent, lifting a 124B student by 2.33 points on SWE-bench Multilingual and 1.60 on SWE-bench Pro.
Details
Coding-agent trajectories are increasingly used as post-training data, but sampling many trajectories independently per issue produces growing redundancy. Feeding earlier attempts to the agent as advice creates a train-test mismatch because that advice is unavailable at inference. SWE-Trail keeps history outside the agent's context and uses it to pick a state where more exploration is worthwhile. There the agent proposes candidate next actions, and SWE-Trail briefly runs each from the same checkpoint to choose which route continues. Training a 124B student initialized from Ling-3.0-flash-base, SWE-Trail improved mean resolution over independent sampling by 2.33 points on SWE-bench Multilingual and 1.60 on SWE-bench Pro. Without history, the same pipeline gained only 0.11 points on Multilingual.