SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training

TL;DR

SWE-Master tackles the challenge of repository-level software engineering by providing an open, end-to-end post-training framework that jointly optimizes data construction, long-horizon supervised fine-tuning, reinforcement learning with real execution feedback, and efficient inference. It introduces LSP-based code navigation to bridge the semantic gap in code understanding, and employs test-time scaling with a high-fidelity simulator (SWE-World) to exploit additional compute without expensive real runs. The approach combines careful trajectory synthesis, rigorous data filtering, and stability-focused RL, achieving state-of-the-art open-source performance on SWE-bench Verified (61.4% with RL and 70.8% under TTS) and demonstrating substantial efficiency gains via LSP tools and continual training. This work enhances reproducibility and accessibility in software-engineering AI, providing a practical blueprint for building and evaluating autonomous code agents and enabling broader adoption in research and development settings.

Abstract

In this technical report, we present SWE-Master, an open-source and fully reproducible post-training framework for building effective software engineering agents. SWE-Master systematically explores the complete agent development pipeline, including teacher-trajectory synthesis and data curation, long-horizon SFT, RL with real execution feedback, and inference framework design. Starting from an open-source base model with limited initial SWE capability, SWE-Master demonstrates how systematical optimization method can elicit strong long-horizon SWE task solving abilities. We evaluate SWE-Master on SWE-bench Verified, a standard benchmark for realistic software engineering tasks. Under identical experimental settings, our approach achieves a resolve rate of 61.4\% with Qwen2.5-Coder-32B, substantially outperforming existing open-source baselines. By further incorporating test-time scaling~(TTS) with LLM-based environment feedback, SWE-Master reaches 70.8\% at TTS@8, demonstrating a strong performance potential. SWE-Master provides a practical and transparent foundation for advancing reproducible research on software engineering agents. The code is available at https://github.com/RUCAIBox/SWE-Master.

Figures (15)

• Figure 1: Performance overview and scaling analysis of SWE-Master. Left: Comparasion of the perference of various open-source foundational models and SWE agents on SWE-bench Verified. Right: Performance of SWE-Master across different training stages and evaluation metrics.
• Figure 2: Distribution of resolve rates and resolved sample counts across interaction turns for SWE-Gym, SWE-smith, SWE-rebench, and R2E-Gym.
• Figure 3: Evolution of interaction turn distributions through sequential filtering stages. From left to right: (1) the initial distribution of successful and failed trajectories; (2) the distribution after applying format and reward constraints; and (3) the final refined distribution following difficulty-based filtering.
• Figure 4: Difficulty distribution across SWE datasets estimated via best-of-n performance. The datasets include SWE-Gym, SWE-smith, SWE-rebench, and R2E-Gym.
• Figure 5: The training dynamics of interaction turns, reward and entropy for SWE-Master RL training.