New benchmarks test whether LLM agents can evolve skills

Several new arXiv papers examine whether LLM agents can improve over time by accumulating skills, memories or experience. ContinualSkillBench finds sequential execution generally improves performance across five domains, but gains vary by model and domain, and in-context learning performs comparably to explicit skill maintenance on average. Related work studies field-aware skill retrieval, streaming evaluations for self-evolving agents, traceable turn memory for agentic reinforcement learning, and LLM-driven optimization of recommendation systems.