Hugging Face Releases SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training
Summary:Hugging Face introduces SWE-Trail, a novel trajectory collection method designed to enhance the efficiency of coding agent training. By leveraging historical trajectories to guide subsequent data collection, SWE-Trail mitigates redundancy issues inherent in independent sampling approaches. Importantly, it keeps the history outside the coding agent's context to prevent train-test mismatches. Experiments demonstrate that SWE-Trail improves mean resolution by 2.33 and 1.60 percentage points on SWE-
Technical Mechanism Analysis
The core mechanism of SWE-Trail lies in using historical trajectories to guide subsequent data collection processes without exposing this history directly to the coding agent. During each new rollout, SWE-Trail analyzes historical trajectories to identify states where additional collection effort may be beneficial. The coding agent then proposes candidate next actions from its regular context, and SWE-Trail executes these actions from the same checkpoint to decide which path to continue. This approach not only reduces redundancy inherent in independent sampling but also avoids exposing historical information as advice to the agent, thus preventing train-test mismatches.
Engineering Trade-offs and Performance
Experiments demonstrate that SWE-Trail performs exceptionally well on the SWE-bench Multilingual and SWE-bench Pro benchmarks. Compared to independent sampling, SWE-Trail improves the mean resolution by 2.33 and 1.60 percentage points on Ling-3.0-flash-base initialized 124B models for Multilingual and Pro, respectively. Even without history, the same pipeline outperforms independent sampling but only improves by 0.11 points on Multilingual. This indicates that the guidance from historical information is crucial for enhancing model performance.
Developer Deployment and Recommendations
For developers, SWE-Trail offers a more efficient and targeted data collection method that significantly reduces redundancy and improves training efficiency. In practice, it is recommended that developers adjust the parameters of SWE-Trail, such as the length and selection strategy of historical trajectories, according to the specific task requirements. Additionally, considering the computational resource requirements of SWE-Trail, developers can further optimize the resource utilization of the training process by combining distributed computing and mixed-precision training techniques.
Conclusion
SWE-Trail demonstrates the great potential of history-guided trajectory collection in enhancing the efficiency of coding agent training. Its advantages in reducing redundancy and avoiding train-test mismatches make it an important tool for future AI model training.
— END —Source: Hugging Face Daily Papers (2026-11-01)
Tags: #Hugging Face #Intelligent Agents #Training Optimization #SWE-Trail
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments