New arXiv papers advance self-evolving agents with skill-based RL

Eleven arXiv papers (July 27–30) target self-evolving LLM agents, proposing methods like FlowEvo, Skill Self-Play, and SERPO that co-evolve reusable skills with reinforcement learning. One quantifies a "regression tax" where added skills can hurt agent performance; others address reward sparsity and rollout efficiency.