AgentPSO: Multi-Agent Particle Swarm Optimization for Evolving LLM Reasoning Skills
A new framework called AgentPSO, inspired by particle swarm optimization, aims to evolve multi-agent reasoning skills in large language models. Unlike existing methods that rely on static agents and inference-time debate or aggregation, AgentPSO treats each agent as a particle-like reasoner with a natural-language skill state and a semantic update direction. Through iterative training, agents update their skills by combining previous velocity, personal-best skill, and global-best skill, improving both individual and collective reasoning performance. The approach addresses vulnerabilities to incorrect peer influence and biased consensus, enabling agents to evolve across tasks. The paper is published on arXiv under identifier 2605.08704.
Key facts
- AgentPSO is a particle-swarm-inspired framework for evolving multi-agent reasoning skills.
- It treats each agent as a particle-like reasoner with a natural-language skill state.
- Velocity is a semantic update direction.
- Agents iteratively move toward stronger skill states.
- Each agent updates its skill by combining previous velocity, personal-best skill, and global-best skill.
- The framework aims to improve both individual and collective reasoning performance.
- It addresses vulnerabilities to incorrect peer influence and biased consensus.
- The paper is published on arXiv with identifier 2605.08704.
Entities
Institutions
- arXiv