Researchers have proposed EVISKILL, an evidence-driven framework for continual skill evolution in LLM agents, according to a paper posted on Hugging Face. The approach targets a persistent problem: agents can accumulate reusable procedural knowledge from interaction experience without updating their underlying model parameters, but existing methods often lose the behavioral evidence and task context that justify each change.
EVISKILL organizes execution observations into Replayable Evidence Cards and synthesizes edits with explicit links to the contexts that support them. Targeted replay then verifies those edits by re-executing them and feeds back corrections where needed.
The framework also addresses a limitation of global validation. A single overall outcome can discard locally supported corrections when a broader revision is rejected, or it may judge an edit prematurely when more experience is needed. EVISKILL instead preserves evidence across epochs and provisionally retains supported edits for further refinement, while global validation governs whether they are incorporated into the final skill.
The authors report experiments on three interactive benchmarks across six LLM backbones, stating that the approach demonstrates effectiveness. Code is available at the project's GitHub repository.