by kjhughes on 7/15/25, 2:03 PM with 1 comments
by alganet on 7/15/25, 2:41 PM
I would prefer if first, the feasibility of a unified integrated theory was proven.
> An important step towards a unified theory of cognition is to build a computational model that can predict and simulate human behaviour in any domain
Why is this step important? Do LLMs qualify as a valid computational model that explains cognition? Which other steps lead to this one?
> Centaur was designed in a data-driven manner by fine-tuning a state-of-the-art large language model
These statements are contradictory. You don't design a large language model, you design it's inference engine. Calling it a "design" implies you had blueprint-like plans for its desired outcome. It's not only a definition argument: language models are fine-tuned by selected input data, and this practice in a research setting raises a red flag.
> We transcribed each of these experiments into natural language, which provides a common format for expressing vastly different experimental paradigms
References for this translation process are from the same authors, and build upon models that are not open-weight. This raises a red flag.
> simplifications were made where appropriate
What was the criteria for determining when a simplification was appropriate or not? I could not find any mention to simplification procedures in the supplementary material.
> Finally, we verified that Centaur fails at predicting non-human behaviour.
Is it failing at predicting non-human behavior, or is it relying on how relatively unknown LLM behavior is?
Let me explain better: if you get participants experienced in exploting LLMs, would Centaur fare differently? This skill is definitely within the realm of human cognition (eg. making an LLM hallucinate). This question is important.