Skill or Skip? Learning Selective Skill Invocation in Agentic Tasks via Dual-Granularity Preference Learning
Summary
SelSkill, a dual-granularity preference-learning framework, addresses the challenge of unhelpful skill invocations in agentic tasks. Existing methods often overlook whether a relevant skill should actually be invoked at a given decision point, which can introduce irrelevant context and disrupt execution. SelSkill formulates skill use as a "skill-or-skip" decision, prioritizing candidate decision points using predictive uncertainty. It constructs controlled invoke-skip preference pairs from shared trajectory prefixes, combining episode-level outcome preferences with step-level invocation preferences. On ALFWorld with Qwen3-8B, SelSkill improved task success by 10.9 percentage points and execution precision by 29.1 percentage points. On BFCL, it boosted task success by 5.7 percentage points and execution precision by 29.5 percentage points. Zero-shot results on Tau-bench and PopQA also demonstrated transferability to new domains.
Key takeaway
For AI Engineers developing agentic systems, focusing on selective skill invocation can significantly enhance task success and execution precision. You should consider implementing preference learning for "skill-or-skip" decisions to prevent unhelpful invocations. This approach avoids introducing irrelevant context and streamlines the agent's execution process, leading to more reliable and effective autonomous agents in complex environments.
Key insights
Learning when to skip a skill invocation is crucial for improving agent performance and precision.
Principles
- Skill use is a "skill-or-skip" decision.
- Combine episode-level and step-level preferences.
- Prioritize decision points using predictive uncertainty.
Method
SelSkill formulates skill use as a skill-or-skip decision, prioritizes decision points via predictive uncertainty, constructs invoke-skip preference pairs from shared trajectory prefixes, and integrates episode-level and step-level preferences.
In practice
- Implement dual-granularity preference learning.
- Use predictive uncertainty for decision point prioritization.
- Construct controlled invoke-skip preference pairs.
Topics
- Agentic AI
- Skill Invocation
- Preference Learning
- Large Language Models
- Task Success
- Execution Precision
- Qwen3-8B
Best for: Research Scientist, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.