It's not. The LLM itself only calculates the probabilities of the next token. Assuming no race conditions in the implementation, this is completely deterministic. The popular LLM inference engine llama.cpp is deterministic. It's the job of the sampler to actually select a token using those probabilities. It can introduce pseudo-randomness if configured to, and in most cases it is configured that way, but there's no requirement to do so, e.g. it could instead always pick the most probable token.
This is a poor conceptualization of how LLMs work. No implementations of models you’re talking to today are just raw autorrgressive predictors, taking the most likely next token. Most are presented with a variety of potential options and choose from the most likely set. A repeated hand and flop would not be played exactly the same in many cases (but a 27o would have a higher likelihood of being played the same way).
>No implementations of models you’re talking to today are just raw autorrgressive predictors, taking the most likely next token.
Set the temperature to zero and that's exactly what you get. The point is the randomness is something applied externally, not a "core concept" for the LLM.
Set the temperature to zero and that's exactly what you get.
In some NN implementations, randomness is actually pretty important to keep the gradients from getting stuck at local minima/maxima. Is that true for LLMs, or is it not something that applies at all?
I'm not sure, hence the question. AFAIK temperature only comes into play at inference time once the distribution is known, but I don't know if there are other places where random numbers are involved.
Eg you tend to randomly shuffle your corpus to train on. If you use drop-out (https://en.wikipedia.org/wiki/Dilution_(neural_networks)) you use randomness. You might also randomly perturb your training data. Lots of other sources of randomness that you might want to try.
The amount of problems where people are choosing a temperature of 0 are negligible though. The reason I chose the wording “implementations of models you’re talking to today” was because in reality this is almost never where people land, and certainly not what any popular commercial surfaces are using (Claude code, any LLM chat interface).
And regardless, turning this into a system that has some notion of strategic consistency or contextual steering seems like a remarkably easy problem. Treating it as one API call in, one deterministic and constrained choice out is wrong.
Counter argument - generating probabilistic tokens (degree of randomness) is core concept for an LLM.