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

In three linesSelSkill, a dual-granularity preference-learning framework, optimizes selective skill invocation in agentic tasks. On ALFWorld with Qwen3-8B: +10.9 pp task success, +29.1 pp execution precision. On BFCL: +5.7 pp task success, +29.5 pp execution precision. Zero-shot transfer to Tau-bench and PopQA.

Summary generated by Claude — human-verified