Researchers caution against attributing human-like reasoning to intermediate tokens in AI models, emphasizing the need for clearer interpretation methods.
Stop Anthropomorphizing Intermediate Tokens As Reasoning/Thinking Traces (2025)
