Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents

TL;DR

This paper introduces guided search strategies for non-serializable environments to improve large language model performance in agentic software engineering tasks, achieving state-of-the-art results.

Contribution

It proposes two novel search strategies guided by learned value functions tailored for non-serializable environments, enhancing success rates of LLMs in complex tasks.

Findings

Double the success rate on SWE-bench Verified with a fine-tuned model

Achieved 40.8% success rate, setting a new state-of-the-art

Transferable techniques improve performance of advanced closed models like GPT-4o

Abstract

Large language models (LLMs) have recently achieved remarkable results in complex multi-step tasks, such as mathematical reasoning and agentic software engineering. However, they often struggle to maintain consistent performance across multiple solution attempts. One effective approach to narrow the gap between average-case and best-case performance is guided test-time search, which explores multiple solution paths to identify the most promising one. Unfortunately, effective search techniques (e.g. MCTS) are often unsuitable for non-serializable RL environments, such as Docker containers, where intermediate environment states cannot be easily saved and restored. We investigate two complementary search strategies applicable to such environments: 1-step lookahead and trajectory selection, both guided by a learned action-value function estimator. On the SWE-bench Verified benchmark, a key testbed for agentic software engineering, we find these methods to double the average success rate of a fine-tuned Qwen-72B model, achieving 40.8%, the new state-of-the-art for open-weights models. Additionally, we show that these techniques are transferable to more advanced closed models, yielding similar improvements with GPT-4o.