Large language models (LLMs) have transformed the software engineering landscape. Recently, numerous LLM-based agents have been developed to address real-world software issue fixing tasks. These agents are typically equipped with various coding tools, enabling them to autonomously generate action trajectories for end-to-end problem-solving. Despite their state-of-the-art performance, these agents still face two primary challenges: (1) Lack of sufficiently high-quality issue descriptions: Inconsistencies between issue descriptions and their corresponding solutions introduce noise and ambiguity that can mislead automated agents, restricting their ability to derive effective solutions. Furthermore, high-quality issue-PR pairs are difficult to acquire at scale, and the available datasets remain relatively limited in size. (2) Uncertainty introduced by bash tools prevents stable training and sufficient exploration: Bash tools serve as the primary interface for code execution and file manipulation in SWE agents, making them integral to virtually all agent operations. However, bash tools introduce substantial uncertainty into the agent’s decision-making process. In reinforcement learning with verifiable rewards (RLVR), the fixed probability ratio clipping mechanism exhibits limited adaptability to such varying uncertainty states. Specifically, when bash tool invocations trigger high-entropy states, excessively restrictive clipping constraints impede both exploration and stable learning, hindering the training efficiency.

To address these challenges, we propose SWE-Fuse, an issue-description-aware training framework that fuses issue-description-guided and issue-free samples for training SWE agents. It consists of two key modules: (1) An issue-free-driven trajectory learning module, which comprises: a multi-step trajectory construction component for generating high-quality multi-turn reasoning-action trajectories, a trajectory data filtering process for ensuring data quality, and issue-free-driven supervised fine-tuning to mitigate potentially misleading issue descriptions while enabling the model to learn step-by-step debugging processes; and (2) An entropy-aware RLVR training module, which adaptively adjusts training dynamics through entropy-driven clipping. It applies relaxed clipping under high entropy to encourage exploration, and stricter clipping under low entropy to ensure training stability. We evaluate SWE-Fuse on the widely studied SWE-bench Verified benchmark shows to demonstrate its effectiveness in solving real-world software problems. Specifically, SWE-Fuse outperforms the best 8B and 32B baselines by 41.9% and 59.3% in solve rate, respectively. Furthermore, integrating SWE-Fuse with test-time scaling (TTS) enables further performance improvements, achieving solve rates of 49.8% and 65.2% under TTS@8 for the 8B and 32B models, respectively.