RSIAgent: Training-Free Memory Lifts Kimi-K3, GLM-5.3 Past GPT-6
TL;DR
- RSIAgent is a training-free multi-agent framework that improves LLM agents by building reusable environment memory rather than updating model weights.
- The paper claims Kimi-K3 and GLM-5.3 with RSIAgent outperform frontier closed-source models including GPT-6 on OSWorld-v2 and Agent's Last Exam.
- The system uses curriculum, actor, and verifier agents in a 'broad-then-deep' exploration strategy that yields frozen memory reusable on downstream tasks.
A training-free multi-agent framework posted to arXiv reports that open-source Kimi-K3 and GLM-5.3, wrapped in a system called RSIAgent, "outperform frontier closed-source models including GPT-6" on OSWorld-v2 and Agent's Last Exam. No model weights are touched.
The mechanism is not new training but new memory. RSIAgent coordinates three specialised agents (curriculum, actor, and verifier) to explore an environment and record what the authors describe as "reusable causal relationships between actions, conditions, and consequences." Exploration follows what they call a "broad-then-deep" pattern: parallel breadth to map environment structure, then focused digs into "hard cases, hidden constraints, boundary conditions, and previously unknown causal dependencies."
The resulting memory, the paper states, "is frozen and can be directly reused for downstream tasks without updating model parameters." That puts the work inside a widening class of agentic scaffolds pitched against fine-tuning as the way to close the gap to frontier systems.
The abstract publishes no per-benchmark scores, no per-model deltas, and no ablation numbers. Two open-source bases beating a frontier closed model on two agentic suites is a strong headline claim, and readers only have the 49-page paper's abstract to test it against.
Originally reported by arxiv.org
Read the original article →Original headline: RSIAgent Paper: Training-Free Recursive Self-Improvement Lifts Kimi K3 and GLM-5.3 Past GPT-6