- the only real comparisons they make are with the parental model, llama 3.7-70b without fine tuning. That tells us what is the added value of fine tuning the dataset but it is hardly state of the art. I guess it should be seen as an indication of how difficult it is to stay afloat in this world when you are in academia and the tech barrier stands billion dollars tall.

- Fig 4a shows Centaur clusters more closely to humans than any other model in a cognitive benchmark (CogBench) but also shows that parental llama cluster closer than claude and openAI thinking models which makes me a bit sceptical of using this measurement at all and reinforces the need for further comparisons.

- the fMRI stuff makes no sense and transforms the paper into a propaganda stunt, IMHO.

- At the end of the paper, the comparison with an "informed" Deepseek-R1 (not shown in data?) shows that a modern reasoning model matches Centaur-performance even without any fine tuning.

The latter point is incredibly interesting in principle but it has nothing to do with the claims of the paper. It basically concludes that a modern reasoning model with CoT can outperform out of the box a "simpler" model that was specifically fine-tuned with a huge dataset of human cognitive behaviours. Bigger claim than the title itself basically IMO.