In this position paper, we argued against the prevalent tendency to anthropomorphize intermediate tokens as reasoning or “thinking”. Anthropomorphization has been a part of AI research [33], and has significantly increased in the era of LLMs [20]. While some anthropomorphization has been harmless metaphors, we argued that viewing intermediate tokens as reasoning traces or “thinking” is actively harmful, because it engenders false trust and capability in these systems, and prevents researchers from understanding or improving how they actually work.
I think it’s a valiant attempt to not over index on what token matching is.. but that ship sailed a long time ago. How is stock market going to go brrrr unless ai robot has a name?
Anyways. We are deep in an investment bubble. Lets see what happens.
Quote Citation: Arizona State University (CS), “STOP ANTHROPOMORPHIZING INTERMEDIATE TOKENS AS REASONING/THINKING TRACES!”, 2025-05-27, https://arxiv.org/pdf/2504.09762