SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training

SWE-Trail. History-Guided Trajectory Collection for Coding Agent Training. By Ziming Li and five others, from the University of Connecticut and Ant Group. Published on the first of November.

Coding agents solve repository issues through long sequences of exploration, editing, and validation. Those trajectories are increasingly used as post-training data, which means the way they are collected shapes what the trained model learns. A common approach samples multiple trajectories independently for each issue, then selects or curates them. The trouble is diminishing returns. As more trajectories are sampled independently, the later ones become increasingly redundant, even though the earlier ones could have guided where further collection effort should be spent. Each new sample starts fresh, as if nothing had been learned about that issue so far.

The natural fix is to show that history to the coding agent as advice. The authors explain why that backfires. Training on trajectories produced with advice can make the model rely on advice, and at test time there is none to rely on. That is a train-test mismatch, built into the data.

Their alternative splits the job in two. History should guide where collection effort is spent, while the coding agent decides what to generate. SWE-Trail is a trajectory collector built on that division. It uses earlier trajectories to guide subsequent collection, while keeping that history outside the coding agent's context altogether.

During a new rollout, SWE-Trail uses the history to identify a state where additional collection effort may be useful. At that state, the coding agent proposes candidate next actions from its ordinary context, the same context it would have with no history at all. SWE-Trail then briefly executes those candidates from the same checkpoint, and uses what happens to decide which route continues. The history steers where the branching happens, and the agent's own proposals supply what happens there. Nothing the agent sees refers to earlier attempts, which is what keeps the resulting data free of advice the model would never get at test time.

Collection approaches are judged by their downstream effect: student models are trained on the trajectories each approach produces, and the students are compared. For a one hundred and twenty-four billion parameter student initialized from Ling-3.0-flash-base, SWE-Trail improves mean resolution over independent sampling by two point three three percentage points on SWE-bench Multilingual. On SWE-bench Pro, the gain is one point six zero percentage points.

A further comparison isolates the role of history. Without it, the same pipeline still produces more resolved trajectories than independent sampling does. But the student trained on them improves by only zero point one one points on Multilingual. More solved trajectories alone bought very little; the history-guided choice of where to spend effort is what moved the result.

The design rests on one separation. Information that helps gather data does not have to be information the model sees. Keeping history on the collector's side lets earlier attempts shape later ones, without teaching the model to expect help it will not have when it is actually deployed on new issues.