Open-SWE-Traces: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents

2026-03-17 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Expert, extended

Summary

Open-SWE-Traces is a new, extensive dataset comprising 207,489 agentic trajectories designed to advance autonomous software engineering. Sourced from 20,000 real-world pull requests via OpenHands and SWE-agent, it covers nine programming languages including Python, Go, and Java. The dataset employs a dual-mode synthesis, using Minimax-M2.5 for "thinking" traces and Qwen3.5-122B for "non-thinking" traces, supporting a "switchable" reasoning framework. Filtered for permissive licenses (MIT, Apache, BSD) from SWE-rebench-V2, Open-SWE-Traces facilitates training models for long-horizon reasoning. Validation through fine-tuning the Qwen3-30B-A3B series yielded a model achieving 61.7% resolve on SWE-bench Verified, 57.1% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro. This resource is crucial for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs.

Key takeaway

For AI Engineers developing autonomous software engineering agents, you should prioritize training with diverse, dual-mode multilingual datasets like Open-SWE-Traces. This approach, incorporating both "thinking" and "non-thinking" trajectories, significantly improves issue resolution across multiple languages and complex tasks. Consider including unresolved task data, as it enhances model robustness. Your models will achieve higher resolve rates and better generalize to real-world scenarios.

Key insights

Dual-mode multilingual distillation with diverse agentic trajectories significantly improves LLM performance in software engineering tasks.

Principles

• Multilingual data boosts cross-lingual transfer.
• Including unresolved trajectories enhances learning.
• Thinking modes improve execution efficiency.

Method

Open-SWE-Traces construction involves repository selection (SWE-rebench v2, permissive licenses, 9 languages), dual-mode trajectory synthesis using MiniMax-M2.5 and Qwen3.5-122B in OpenHands/SWE-agent, and multi-stage quality filtering including AST-based "git hacking" detection.

In practice

• Train agents with both "thinking" and "non-thinking" traces.
• Incorporate multilingual data for broader task resolution.
• Include unresolved task trajectories in training data.

Topics

• Software Engineering Agents
• LLM Fine-tuning
• Multilingual Datasets
• Dual-Mode Reasoning
• SWE-bench Benchmarks
• Trajectory Distillation

Code references

• multi-swe-bench/MopenHands
• multi-swe-bench/MSWE-agent
• zhenyuhe00/SWE-Swiss

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• MADE: Beyond Scoring via a Multilingual Agentic Diagnosing Engine for Fine-Grained Evaluation Insights — Multilingual LLM evaluation requires fine-grained, agentic diagnosis to transform scores into actionable insights, addressing cultural and deployment realities.
• Drawing on Memory: Dual-Trace Encoding Improves Cross-Session Recall in LLM Agents — Dual-trace memory encoding enhances LLM agent recall by adding contextual "scene traces" to factual memories.
• Beyond SWE-Bench Pro - Where do Agents go from Here? — New benchmarks are crucial for advancing AI coding agents beyond simple issue resolution to complex system understanding and improvement.
• Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills — Trace2Skill distills broad agent experience into transferable, declarative skills via parallel analysis and inductive consolidation.
• Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models — Aligning multi-turn LLM behavior with full-context teacher responses mitigates self-anchored drift.
• Most Influential ArXiv (Software Engineering) Papers (2026-04 Version) — LLMs and AI agents are rapidly transforming software engineering, driving innovation in code generation, testing, and repair.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.