New preprint! 🤖🧠
The cost of thinking is similar between large reasoning models and humans
👉 osf.io/preprints/psyarxiv/m2…
w/ Ferdinando D'Elia, @AndrewLampinen, and @ev_fedorenko (1/6)
Jul 25, 2025 · 5:32 PM UTC
A central goal in CogSci is to develop models that explain human reasoning across diverse tasks, from arithmetic to logic to relational inference.
Large reasoning models can now solve many of these problems. But does their reasoning process reflect how humans think? (2/6)
One key aspect is processing cost, reflected in reaction times: how long does it take to solve a problem?
We compared human RTs to the length of DeepSeek-R1's reasoning traces across five domains: arithmetic (numeric & verbal), logic, relational reasoning, and the ARC task (3/7)
We found that the number of reasoning tokens generated by the model reliably correlates with human RTs within each task (mean r = 0.62, all ps < .001). (4/7)
And: this alignment is stronger for R1 than for its base model (DeepSeek-V3), which isn't trained to reason step-by-step. Training the model to reason improves its match to human processing cost. (5/7)
Token count also captures differences across tasks. Avg. token count predicts avg. RT across domains (r = 0.98, left), and even item-level RTs across all tasks (r = 0.95 (!!), right).
A single effort metric scales with human processing cost across domains. (6/7)
