Confession: I haven't read the paper. But any mention of LLM reasoning ability ought to address the obvious confound: the LLM is trained on examples of deducti

Confession: I haven't read the paper. But any mention of LLM reasoning ability ought to address the obvious confound: the LLM is trained on examples of deductive reasoning, inductive reasoning, abductive reasoning, SAT-solver reasoning, geniuses' musings, etc. If they replicate one of those examples, then should that be called "reasoning" of any sort or not? Regurgitating those examples may even involve some generalization, if the original topics of an example are swapped out (perhaps by a nearby topic in latent space). Given that it appears they're training and testing on synthetic problems, this objection probably does not apply to their actual results. But given the fuzziness it creates for the definition of "reasoning" of any sort, I would have expected some working definition of reasoning in the paper's abstract. Training on Moby Dick and thus being able to regurgitate text from Moby Dick does not mean the LLM is capable of writing a new Moby Dick-like book. (Thankfully; one is more than enough!)