Research map / Multi-agent systems and games / LLM agents
SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents
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 SE-Agent, a self-evolution framework that iteratively revises, recombines, and refines multi-step agent trajectories to improve SWE-bench Verified resolution rates across five LLMs.
Problem
LLM agents' multi-step interaction trajectories are underexploited; MCTS and independent sampling yield homogeneous, redundant trajectories and suboptimal outcomes on complex tasks such as GitHub issue resolution.
Approach
Treats pilot trajectories as an improvement operator: generates diverse initial trajectories via multi-planning and mutation, then evolves them with genetic-style revision (self-reflection), recombination (crossover, transfer, restructuring), and refinement (multi-dimensional reward and elite selection) over several iterations, built on SWE-Agent.
Key results
- SE-Agent with Claude-3.7-Sonnet reaches 61.2% Pass@1 on SWE-bench Verified vs 40.6% for SWE-Agent (+51% relative).
- With Claude-4-Sonnet on the latest SWE-Agent, SE-Agent reports 80.0% resolved versus a 66.6% baseline.
- Relative gains up to 112% (Llama-3.1-70b-Instruct) over SWE-Agent; SE-Agent uniquely solves 12 instances no other compared model solves.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SE-Agent instead of SWE-Search | DeepSeek-V3-0324 · SWE-bench Verified | Pass@1: 39.4% -> 54.8% | Helped |
| Use SE-Agent instead of SWE-Search | DeepSeek-V3-0324 · SWE-bench Verified | Pass@5: 41.8% -> 58.4% | Helped |
| Use SE-Agent instead of SWE-Search | Qwen-2.5-72b-Instruct · SWE-bench Verified | Pass@1: 23.4% -> 38.8% | Helped |
| Use SE-Agent instead of SWE-Search | Qwen-2.5-72b-Instruct · SWE-bench Verified | Pass@5: 26.2% -> 42.4% | Helped |
| Use SE-Agent instead of SWE-Search | Llama-3.1-70b-Instruct · SWE-bench Verified | Pass@1: 21.8% -> 32.6% | Helped |
| Use SE-Agent instead of SWE-Search | Llama-3.1-70b-Instruct · SWE-bench Verified | Pass@5: 23.6% -> 35.2% | Helped |
| Use SE-Agent instead of SWE-Search | GPT-4o · SWE-bench Verified | Pass@1: 32.6% -> 40.4% | Helped |
| Use SE-Agent instead of SWE-Search | GPT-4o · SWE-bench Verified | Pass@5: 35.8% -> 44.8% | Helped |
| Use SE-Agent instead of SWE-Search | Claude-3.7-Sonnet · SWE-bench Verified | Pass@1: 47.4% -> 61.2% | Helped |
| Use SE-Agent instead of SWE-Search | Claude-3.7-Sonnet · SWE-bench Verified | Pass@5: 50.6% -> 63.6% | Helped |
| add revision operation (ablation) | SE-Agent · SWE-bench Verified | Pass@1 | Helped |
| add recombination operation (ablation) | SE-Agent · SWE-bench Verified | Pass@1 | Helped |
Surprising findings
- Repeated sampling under identical configurations yields homogeneous outcomes, motivating trajectory-level rather than sampling-level diversity.
Limitations
- Gains evaluated only on SWE-bench Verified; generalization to other multi-step tasks is not demonstrated.
- Relies on LLM-based evaluation (ExpertEval) and handcrafted reward weights, with no reported sensitivity analysis of weights.
- Multiple LLM calls per iteration increase API cost relative to baselines.
Experimental setup
Datasets: SWE-bench Verified
Benchmarks: SWE-bench Verified
Metrics: Pass@1 Pass@5 resolution rate
Baselines: SWE-Agent SWE-Search
Builds on: ReAct SWE-Agent Monte Carlo Tree Search Reflexion
Closest papers by meaning
- Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.90)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.89)
- PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.89)
- MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs (NeurIPS 2026 · similarity 0.89)
- AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.89)
- PathWise: Planning through World Model for Automated Heuristic Design via Self-Evolving LLMs (ICML 2026 · similarity 0.89)
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.