Rsiagent

We released RSIAgent, a framework for recursive self-improvement in new environments. Curriculum, actor, and verifier agents autonomously explore, verify outcomes, and consolidate experience into memory that guides subsequent rounds. Built on open-source GLM-5.3 and Kimi-K3, it outperforms GPT-6 Astra in partial-credit scores on OSWorld 2.0 and Agents’ Last Exam. Hugging Face Daily Papers #6 (Sep 15, 2026); Headline feature in Synced (机器之心), an AI news outlet. See the code and project page.