Research map / Multi-agent systems and games / LLM agents
SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SWE-Search, a multi-agent framework combining MCTS with an LLM hybrid value function (numeric score plus natural-language feedback) and a debate-based discriminator, improving SWE-bench Lite resolve rates by about 23% relative across five models.
Problem
LLM software agents follow linear, sequential processes that prevent backtracking and exploration of alternative solutions, limiting performance on repository-level software engineering tasks.
Approach
Applies MCTS to a moatless-tools style software agent with a git-based commit tree for reverting states, using an LLM value agent that outputs both a numeric value and a natural-language explanation fed back into re-expansion (hindsight feedback), plus a multi-agent debate discriminator to select the final patch. The selection function is heuristic (UCT with depth bonus and penalty) rather than LL
Key results
- SWE-Search with GPT-4o resolves 31.0% on SWE-bench Lite vs 25.7% for Moatless-Adapted (+17% relative).
- Averaged over five models, SWE-Search gives a +23% relative improvement over Moatless-Adapted.
- Discriminator selects the correct solution 84% of the time vs 73% for the value function.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MCTS search and iterative value feedback | Moatless-Adapted with GPT-4o · SWE-bench Lite | resolve rate (Pass@1): 25.7 -> 31.0 | Helped |
| Add MCTS search and iterative value feedback | Moatless-Adapted with GPT-4o-mini · SWE-bench Lite | resolve rate (Pass@1): 13.0 -> 17.0 | Helped |
| Add MCTS search and iterative value feedback | Moatless-Adapted with Qwen2.5-72B-Instruct · SWE-bench Lite | resolve rate (Pass@1): 18.0 -> 24.7 | Helped |
| Add MCTS search and iterative value feedback | Moatless-Adapted with DeepSeek-V2.5 · SWE-bench Lite | resolve rate (Pass@1): 16.3 -> 21.0 | Helped |
| Add MCTS search and iterative value feedback | Moatless-Adapted with Llama-3.1-70B-Instruct · SWE-bench Lite | resolve rate (Pass@1): 13.6 -> 17.7 | Helped |
| Allow Plan states to transition to any state type (ablation) | Moatless-Adapted · SWE-bench Lite | resolve rate (Pass@1): 24.3 -> 25.7 | Mixed |
| Add MCTS and search-and-feedback to flexible state transitions | Moatless-Adapted · SWE-bench Lite | resolve rate (Pass@1): 25.7 -> 31.0 | Helped |
| Add a structured multi-agent discriminator debate for final solution selection (ablation) | SWE-Search · SWE-bench Lite | correct final-solution selection: 73% -> 84% | Helped |
| Use state-specific value prompts for Searching, Planning, and Editing (ablation) | SWE-Search Value Agent · SWE-bench Lite | value assessment of effective states | Helped |
Surprising findings
- Flexible Plan-state transitions in Moatless-Adapted yield only a 1.4% gain over the original moatless-tools, suggesting unconstrained flexibility can trap agents in loops.
- 33 issues were solved by other models but not by GPT-4o, indicating model diversity matters.
Limitations
- Search budget is conservative (100 iterations) so scaling behavior at larger compute is not established.
- Value function accuracy (73%) leaves room for improvement in identifying correct solutions.
- Flexible state transitions can cause infinite loops without high-level control.
Experimental setup
Datasets: SWE-bench Lite
Benchmarks: SWE-bench Lite
Metrics: Pass@1 resolve rate Pass@5 relative improvement
Baselines: Moatless-v0.0.2 Moatless-Adapted SWE-agent Agentless AutoCodeRover OpenDevin
Builds on: moatless-tools SWE-bench AlphaZero UCT
Closest papers by meaning
- Discovering Hierarchical Software Engineering Agents via Bandit Optimization (ICLR 2026 · similarity 0.91)
- Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents (ICML 2025 · similarity 0.91)
- SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents (NeurIPS 2025 · similarity 0.90)
- SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering (NeurIPS 2024 · similarity 0.89)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.89)
- FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.