I think the article is part right and part wrong.

The right part: autoregressive LLMs are indeed generating “probabilities” (scare quotes very much intentional). During pre-training and any SFT steps, those probabilities are nudged toward the probabilities, over the training distribution, of the next token conditioned on the previous tokens. (This is an explicit property of most training recipes: KL divergence is a “proper scoring function”.)

So if you prompt with “Paris is a city in ”, the next token probabilities estimate the probabilities over the input distribution that the next token in the sentence is the first token of France or of something else.

But there are huge caveats:

1. That is not at all the same thing as the probability that Paris is France under any distribution that you care about (the population of the various Parises, for example).

2. None of this necessarily usefully applies to RL or, as the article discusses, tool calling. The output probability of a tool call is not some Platonic idea of a probability that the input is worthy of a tool call. It’s a the result of a training process that tried to teach the model to be useful and to achieve its goals.

3. I suspect that reasoning makes this all much worse. Suppose that you prompt with “a help desk user with IP=a.b.c.d says they’re ‘in Paris’. What country are they in?” The model has been trained to generate a reasoning trace, which may well start with “let me think of where Paris could be. It could be in France or in Texas etc. The user was speaking English…” See the problem? The model is reasoning well, but it reasoned “France” before “Texas”, so the logprob for France was probably higher than “Texas”. At the end of the reasoning trade there will be an answer, but the logprobs for that answer are, at best, some representation of the probabilities of the answer conditioned on the sampled reasoning trace. And that is not the probability distribution that a Jev user wants.