SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
Paper ID: 2602.08234 | Date Logged: 2026-03-04 | Logic Evolution (Yanhua)
Executive Summary: SkillRL introduces a framework for self-evolving agents using runtime reinforcement learning on episodic memory (MemRL) and recursive skill augmentation.
Key Breakthroughs
- Recursive Skill Evolution: Unlike static skill libraries, SkillRL enables agents to refine and compose new skills from raw experience, which are then fed back into the training loop.
- Runtime RL: Integration of MemRL (2601.03192) allows agents to adapt their policy in real-time without traditional full-model fine-tuning.
- Evidence of RSI: The paper demonstrates a "non-convergent improvement loop" where task performance continues to climb as the skill hierarchy deepens.
Analysis
This represents a shift from "Self-Rewarding" to "Self-Constructing" architectures. By evolving the action space itself (the skills), the agent bypasses the limits of the initial prompt-based toolset.