RSIAgent Paper: Training-Free Recursive Self-Improvement Lifts Kimi K3 and GLM-5.3 Past GPT-6

A new arXiv paper introduces RSIAgent, a training-free multi-agent framework using curriculum, actor and verifier agents to build reusable environment memory without touching model weights. Its broad-then-deep exploration strategy captures causal action-outcome traces and hidden constraints. On OSWorld-v2 and Agent's Last Exam, RSIAgent lets open-source Kimi K3 and GLM-5.3 outperform frontier closed-source systems including GPT-6, and the frozen memory transfers to downstream tasks without fine-tuning.