SWE-Replay: Efficient Test-Time Scaling for Software Engineering Agents

Test-time scaling has been widely adopted to enhance the capabilities of Large Language Model (LLM) agents in software engineering (SWE) tasks. However, the standard approach of repeatedly sampling trajectories from scratch is computationally expensive. While recent methods have attempted to mitigate costs using specialized value agents, they can suffer from model miscalibration and fail to generalize to modern agents that synthesize custom bash scripts as tools. In this paper, we introduce SWE-Replay, the first efficient and generalizable test-time scaling technique for modern agents that does not rely on potentially noisy value estimates. SWE-Replay optimizes the efficiency of the scaling process by recycling trajectories from prior trials, dynamically choosing to either explore from scratch or exploit archived experience by branching at critical intermediate steps. This selection of intermediate steps is driven by the potential and reasoning significance of repository exploration, rather than external LLM-based quality estimates. Our evaluation shows that SWE-Replay reduces the computational cost of test-time scaling by up to 17.4%, while maintaining comparable performance (with slight improvements of up to 3.8%), demonstrating that effective scaling can be achieved without proportional cost increases. Further evaluation on SWE-Bench Pro and Multilingual validates the generalizability of SWE-Replay, establishing it as a robust foundation for efficient test-time scaling of software engineering agents.