Tag
This paper introduces a model-adaptive definition of tool necessity and finds a 26-54% mismatch between LLMs' internal recognition that a tool is needed and their actual tool-call actions, concentrated in the cognition-to-action transition. It reveals a 'knowing-doing gap' where the model often knows it should call a tool but fails to do so due to late-layer geometry rotating the signal nearly orthogonal to the action.
Skill-RAG is a failure-aware RAG framework that uses hidden-state probing and skill routing to diagnose and correct query-evidence misalignment in retrieval-augmented generation. The approach detects retrieval failures and selectively applies targeted skills (query rewriting, question decomposition, evidence focusing) to improve accuracy on hard cases and out-of-distribution datasets.