can LLM reason

I’ve spent some time teaching and even training large language models. The question of whether LLMs can actually reason still comes up often. My answer remains essentially the same: no — but with a grain of salt.

A couple of years ago, my stance was straightforward. LLMs are fundamentally massive probabilistic models, designed to predict the most likely next token. Their outputs are further refined through reinforcement learning, but that doesn’t equate to genuine reasoning.

That said, recent advances like chain-of-thought prompting have shifted the conversation. By encouraging step-by-step problem solving, LLMs now produce more accurate answers. Is this true reasoning? Strictly speaking, no — the model is still operating on probabilities. But chain-of-thought helps steer the model through intermediate steps, reducing error along the way. Each step becomes new input in the autoregressive process, which improves overall accuracy.

So, still no reasoning? Maybe not. The “grain of salt” comes from the fact that humans, too, often reason better when we “think out loud.” In that sense, chain-of-thought bears some resemblance to how people structure their reasoning — and that’s why I keep a small reservation.