Researchers introduced RSIAgent, a multi-agent framework for recursive self-improvement. It uses curriculum, actor, and verifier agents to explore environments and build reusable memory. The system employs a broad-then-deep strategy to uncover diverse structures and hidden constraints. Memory is frozen and reused for downstream tasks without updating model parameters. Experiments show RSIAgent improves open-source models like Kimi-K3 and GLM-5.3. These models outperform frontier closed-source models such as GPT-6 on OSWorld-v2 and Agent’s Last Exam. The framework enables training-free adaptation to new environments. It autonomously discovers causal relationships between actions, conditions, and outcomes. The paper presents RSIAgent as a general solution for agent-level self-improvement.
What changed
Digital agents relied on additional training data and human assistance for adaptation to new environments.
What it unlocks
Autonomous discovery and reuse of environment-specific causal knowledge without model parameter updates.
- Kimi-K3 and GLM-5.3 outperform GPT-6 Astra
- OSWorld-v2 and Agent’s Last Exam benchmarks
- Broad-then-deep exploration strategy
Sources