SkillApt: Learning When to Activate Agent Skills from Counterfactual Evidence
3.60T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.26863.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummarySkillApt is a post-retrieval activation framework for LLM agents that decides whether a retrieved skill should be loaded into context. It uses counterfactual WITH/WITHOUT execution evidence to make LOAD/ABSTAIN decisions per candidate skill. On SRA-Bench it matched BM25 Top-1 accuracy (0.838) while cutting activation from 100% to 31.5% and mean token usage by 74.3%.
Why it mattersSeparates skill retrieval from skill activation as distinct decisions, with measured token savings. Useful framing for anyone designing agent context-loading strategies.
Cited by
No citations on record.