This is a long paper (50 pages!) by a fellow Greek and I want to take my time and give it the attention it seems to deserve. However, I have to say I'm not sure if such work is really well, necessary.
Here's the thing. The paper, like others, is contributing to the literature around the hypothesis that LLMs can reason. There have been articles both supporting and rejecting the hypothesis, and this one claims it's false.
But, in science, we don't start with a hypothesis. We start with some observations, and then we make up a hypothesis to try and explain the observations. Then we try to reject our hypothesis with more observations. What are the observations that led to the hypothesis that LLMs can reason?
It's one observation really: that LLMs can geneate text that looks like the result of reasoning. There exists a much simpler explanation of this observation, than the hypothesis that LLMs can reason. Namely, LLMs are trained to generate text similar to text generated by humans, who (we assume) can reason. If an LLM is good at that job, then obviously at some point it will generate text that looks like the result of reasoning. The ability to reason is not necessary.
If we have this simpler explanation, there's no reason to reach for the more complex one, that needs more assumptions.
And remember kids: if you multiply entities beyond necessity, out comes the Macco Man and shaves your head with his R A Z O O O O O R R!!!
So don't do that. Assume the simplest explanation until such time as it is untenable.
Here's the thing. The paper, like others, is contributing to the literature around the hypothesis that LLMs can reason. There have been articles both supporting and rejecting the hypothesis, and this one claims it's false.
But, in science, we don't start with a hypothesis. We start with some observations, and then we make up a hypothesis to try and explain the observations. Then we try to reject our hypothesis with more observations. What are the observations that led to the hypothesis that LLMs can reason?
It's one observation really: that LLMs can geneate text that looks like the result of reasoning. There exists a much simpler explanation of this observation, than the hypothesis that LLMs can reason. Namely, LLMs are trained to generate text similar to text generated by humans, who (we assume) can reason. If an LLM is good at that job, then obviously at some point it will generate text that looks like the result of reasoning. The ability to reason is not necessary.
If we have this simpler explanation, there's no reason to reach for the more complex one, that needs more assumptions.
And remember kids: if you multiply entities beyond necessity, out comes the Macco Man and shaves your head with his R A Z O O O O O R R!!!
So don't do that. Assume the simplest explanation until such time as it is untenable.