Measuring when AI models learn and retire skills reliably
Beyond Gain: Measuring Skill Internalization for Agent Skill Retirement
Artificial Intelligence
Summary
When AI models use skills from a shared hub, some skills become unnecessary as models learn to do tasks on their own. The authors show that just checking if performance improves with a skill (gain) misses whether a model truly internalized that skill. They propose a new way to test if an AI really has learned a skill by withholding it and testing on related tasks, making sure the skill truly matters for other models. Their method more reliably detects when skills can be safely removed for each model, helping manage large skill hubs.
What this means in practice
- •For ai system maintainers: Improve managing shared skill libraries by reliably identifying which skills a model no longer needs to reduce unnecessary skill use.
- •For ai developers: Gain more accurate feedback on model progress when training by testing skill internalization rather than only measuring performance gain.
Authors
Yanze Wang, Pengfei Yao, Tianyi Sun, Hongda Chen, Yunyun Han, Jun Sun, Chuanrui Hu, Yafeng Deng
Abstract
LLM agents retrieve skills from public hubs that hold up to about 100,000 entries and serve many models. As models improve, a hub must find the skills a given model has internalized, meaning it can do without them, and retire them for that model, yet internalization has not been measured with known reliability. Published rules that compare runs with and without a skill retire on the gain alone. The gain is near zero both for an internalized skill and when neither the model nor the skill delivers the capability, a state we call uncovered. We instead withhold the skill, score the model on a task synthesized from it, and trust a high unaided score only when another model gains from the skill, since nearly half of high scores come from tasks no model benefits from. Under four models, a published gain rule applied with one task per skill removes over half of the uncovered skills. On SkillsBench, 28 of its 31 retirements rest on probes no run passed in full, while retiring on the unaided score costs no detectable reward. Unaided verdicts reproduce across tasks more often than the gain rule's (Cohen's kappa 0.42 against 0.23) but differ across models, so a hub must read them per model. We release 1,251 skills with tasks and item scores under four models.