$ cat wiki/papers/2026/2611.02074-swe-trail.md
SWE-Trail — using history without teaching dependence on it
TL;DR
SWE-Trail uses prior coding trajectories to decide where to collect more experience, while keeping that history outside the coding agent's context. The authors evaluate the resulting data by training student models, rather than treating more solved collection runs as sufficient evidence. (source)
Authors & Org
Ziming Li, Lu Yu, Jiatan Huang, Qing Cui, Jun Zhou, Chuxu Zhang. Affiliations are not established by the captured abstract record. First submitted 2026-11-01. (source)
Method
Earlier trajectories identify a state worth exploring during a new rollout. The coding agent proposes candidate next actions from its ordinary context; the collector briefly executes them from the same checkpoint to choose which route continues. (source)
Keeping history outside the agent's context is intended to avoid dependence on advice unavailable at test time. The intervention changes collection decisions rather than adding historical advice to the student's deployment input. (source)
Results
For a 124B student initialized from Ling-3.0-flash-base, mean resolution improves over independent sampling by 2.33 percentage points on SWE-bench Multilingual and 1.60 percentage points on SWE-bench Pro. (source)
Without history, the same pipeline collects more resolved trajectories than independent sampling but improves Multilingual by only 0.11 points. These are author-reported differences, not absolute benchmark success rates. (source)
Significance
For Agents (LLM Agents) and Agentic Reinforcement Learning, the useful distinction is between improving the collection procedure and exposing the policy to extra information. Interpretation: downstream student performance tests whether collected experience is useful, beyond the number of solved training rollouts. (source)
Open Questions
The abstract does not quantify total collection compute, equal-budget comparisons, or gains across other student scales and deployment domains. Full paper and code were not reviewed, and independent reproduction is not established here. (source)