AgentPSO: Evolving Agent Reasoning Skill via Multi-agent Particle Swarm Optimization

Researchers have introduced AgentPSO, a novel framework designed to enhance the reasoning capabilities of large language models through multi-agent particle swarm optimization. Unlike existing methods that rely on static agents and inference-time debate, which are prone to biased consensus, AgentPSO treats each agent as a particle with evolving natural-language skills. The system iteratively updates agent states by combining previous velocities, personal bests, global bests, and self-reflective directions derived from peer trajectories. This approach allows agents to learn reusable reasoning behaviors without modifying the backbone model's parameters. Experimental results on mathematical and general reasoning benchmarks demonstrate that AgentPSO outperforms static single-agent skills and test-time-only multi-agent baselines. Furthermore, the evolved skills show strong transferability across different benchmarks and backbone models, indicating the capture of generalizable reasoning procedures rather than benchmark-specific optimizations. The code for this innovative AI framework has been open-sourced, offering a new pathway for improving collective intelligence in artificial systems through dynamic skill evolution.