Skip to content
AI HOT (Curated Pool)

Study on the Cognition–Action Disconnect in Tool-Using Agents

工具使用代理认知与行动脱节机制研究

An interpretability paper studies tool-using agents and finds models often recognize when to call a tool but fail to act, with a cognition-to-action mismatch rate of 26%–54%.

Why it matters: A paper on tool-calling agent interpretability finds models recognize a tool is needed but fail to call it in 26% to 54% of cases, with internal signals turning near-orthogonal to the action in later layers.

Read the original ↗Export Markdown