SKILLER addresses the model-mismatch problem, where reusable skills designed around powerful frontier models fail when deployed on smaller, cheaper language models. It uses a frontier model as both actor and critic to iteratively refine task-specific skills through natural-language feedback, treating a small model’s own agent loop as an interactive environment and propagating optimization signals entirely as text rather than gradients. Across five benchmarks, SKILLER outperformed existing skill-generation methods and allowed a small model (Qwen3.5-4B) to surpass the task-specific performance of larger, unoptimized models at a fraction of the inference cost.
