RSIAgent: Zero Training, Just Environment Exploration Beats GPT-6 on Agent Benchmarks

大模型之路 · wechat · 2026-09-24

RSIAgent (arXiv, Sep 14) is a training-free agent framework: it explores unfamiliar environments first — broad parallel exploration followed by targeted deep-dives, coordinated by a curriculum agent, an executor, and a verifier — then freezes an environment-specific memory for downstream tasks. The authors claim this lets Kimi-K3 and GLM-5.3 beat GPT-6 on OSWorld-v2 and Agent's Last Exam (self-run comparison, not third-party). The core thesis: agent capability gaps often lie in environment knowledge, not model weights, shifting the lever from compute to engineering. Best suited to stable interfaces with diverse tasks; frozen maps go stale under frequent UI releases and need periodic re-exploration.

Related event: RSIAgent: Training-Free Self-Improvement Beats GPT-6 with Open Models(3 posts)→

Original post →

More from coding & agent

coding & agent channel →