Critique of Intermediate Token Generation as Human-Like Reasoning
August 19, 2026
The paper argues that intermediate token generation (ITG) should not be interpreted as human-like reasoning or thinking traces. It suggests that treating these tokens as interpretable windows into a model's cognitive process is a category error that leads to false anthropomorphism.
HOW THIS AFFECTS YOU
●
researcherYou should reconsider how you evaluate and label chain-of-thought outputs in your benchmarks.