SelSkill learns when an agent should invoke a reusable skill and when it should skip it. It combines episode-level and step-level preference learning with controlled invoke-or-skip comparisons to improve selective skill use in agentic tasks.

Publication

Accepted to Findings of EMNLP 2026 and published in Findings of the Association for Computational Linguistics: EMNLP 2026 (October 2026).

ACL Anthology ID: 2026.findings-emnlp.222. Pages: 4665–4681. DOI: 10.18653/2026.findings-emnlp.222.

Overview

Reusable skills can help agents, but indiscriminate invocation can be counterproductive. SelSkill prioritizes uncertain decision points and constructs invoke-or-skip preference pairs that share the same trajectory prefix. It combines episode-level outcomes with step-level signals to train selective skill invocation.

Evaluation

The paper evaluates selective skill invocation on ALFWorld and BFCL, with additional zero-shot transfer experiments on Tau-bench and PopQA.

Resources