Language‑model agents improve by turning execution experience into reusable external skills. Repeated skill updates, however, constitute a learning process of their own: locally useful edits can accumulate into redundant or task‑specific instructions, and new updates may disrupt previously working behavior. We term this phenomenon skill‑evolution overfitting and introduce SkillEvoReg, a general regularization framework inspired by anti‑overfitting techniques in neural‑network training. SkillEvoReg combines training‑time skill dropout (perturbing update generation), complexity‑aware local regularization (limiting unnecessary structural growth), and causal counterexample validation (CCV, providing targeted behavioral checks for candidate regressions). We instantiate the framework across heterogeneous skill‑evolution systems while preserving each system’s native skill evolver and task evaluator. Experiments on SkillOpt, SkillEvolBench, and ContinualSkillBench show that SkillEvoReg consistently curtails skill‑state growth, retains competitive downstream performance, improves several transfer and later‑stage evolution outcomes, and uncovers update‑level regressions that structural metrics alone miss. These results suggest that explicit regularization is a useful complement to increasingly capable skill updaters.
Review