Title: Skill or Skip? Learning Selective Skill Invocation in Agentic Tasks via Dual-Granularity Preference Learning

Abstract:

Callable procedural modules known as agent skills offer reusable knowledge and execution policies essential for navigating complex agentic tasks. While current research primarily concentrates on identifying relevant skills or enhancing their internal capabilities, it often neglects a critical question: should a relevant skill actually be invoked at a specific decision point? Invoking a skill when it is not needed can introduce irrelevant context, potentially derailing an otherwise successful execution process. To resolve this challenge, we introduce SelSkill, a framework for selective skill invocation grounded in dual-granularity preference learning.

SelSkill reframes skill utilization as a binary choice: to invoke or to skip. The method leverages predictive uncertainty to identify and prioritize candidate decision points. By analyzing shared trajectory prefixes, it constructs controlled preference pairs that contrast invocation against skipping. Furthermore, the framework integrates episode-level outcome preferences with step-level invocation preferences, thereby capturing both the global quality of the trajectory and the local efficacy of specific skill calls.

Empirical evaluations demonstrate significant gains. On the ALFWorld benchmark using Qwen3-8B, SelSkill boosts task success rates by 10.9 percentage points and execution precision by 29.1 percentage points. Similarly, on BFCL, it increases task success by 5.7 percentage points and execution precision by 29.5 percentage points. Additionally, zero-shot tests on Tau-bench and PopQA indicate that the learned invocation policy generalizes effectively to new domains featuring previously unseen skills.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC