Why does this alignment emerge? There are similarities in how reasoning models and humans learn: first by observing worked examples (pretraining), then by practicing with feedback (RL). In the end, just like humans, they allocate more effort to harder problems. (6/6)

Token count also captures differences across tasks. Avg. token count predicts avg. RT across domains (r = 0.97, left), and even item-level RTs across all tasks (r = 0.92 (!!), right). (5/6)

We found that the number of reasoning tokens generated by the model reliably correlates with human RTs within each task (mean r = 0.57, all ps < .001). (4/6)

Large reasoning models can solve many reasoning problems, but do their computations reflect how humans think?
We compared human RTs to DeepSeek-R1’s CoT length across seven tasks: arithmetic (numeric & verbal), logic (syllogisms & ALE), relational reasoning, intuitive reasoning, and ARC (3/6)

Neural networks are powerful in-silico models for studying cognition: LLMs and CNNs already capture key behaviors in language and vision. But can they also capture the cognitive demands of human reasoning? (2/6)

In conclusion, our results show that (1) LMs are broadly applicable models of the human language system across languages, and (2) there is a shared component in the processing of different languages (14/14)

Languages greatly vary in form, but there is a massive overlap in the concepts they can express. We speculate that this shared meaning space is responsible for successful encoding transfer, but we’ll look more into this in future work (13/)

What supports the transfer of the encoding models? Form and meaning are two promising candidates. However, form-based (phonological, phonetic, syntactic) language similarity does not predict transfer performance (12/)