On decision models

There’s a lot of fuss about decision models lately. I thought that what I’m going to say was obvious to everyone. But it doesn’t seem to be the case, so here we are.

In the beginning there was Jev

Well, no. Actually Jev is not the first decision model. It just got a lot of visibility.

Experiments with LLM-derived reward models were published back in 2020! Researchers started with a pretrained language model, fine-tuned it, and added a scalar-output head to output score candidate summaries. The scoring model did not generate an answer. It was a specialized evaluator, not a general-purpose interface for arbitrary decision questions.

Other relevant examples, from a quick search online: SayCan (2022), Llama Guard (December 2023), RouteLLM (2024), GLiClass (2024–2025). GLiClass in particular is a generalist zero-shot classification with labels supplied at inference time, evaluating the text and candidate labels in one forward pass, without autoregressive answer generation. It’s based on encoder language models such as DeBERTa and ModernBERT, rather than a decoder-style chat LLM. Apart from this is essentially the same idea as Jev.

So what is a decision model?

An LLM. Really. Or at least, most of the models we talk about are basically LLMs that don’t do an autoregressive generation but instead a single forward pass that outputs classification scores. There are various ways of making one, for simplicity let’s say it’s an LLM with the very last part of the network swapped out. But still mostly the original LLM. You could use any other ML classification model if you train it for that use case, I’d say it qualifies as a decision model.

Why are they relevant?

The importance of decision models, I think, is not in any kind of novelty. The real impact is that they made more clear one fundamental fact:

We don’t train LLMs because we want just a model to predict text.
We want LLMs to be able to reason, so we train them to predict text.
Because in order to predict text, they’re forced to also learn to reason.

Next token prediction is a pretext target used to actually train models to reason (and to embed a lot of information, in most cases). And decision models show it off nicely: they perform simple tasks using that encoded knowledge and capabilities, without ever predicting a single token.

Decision models are generalist.

They can act as generalist classifiers, which is extremely handy in many cases. They’re going to perform worse that most specifically designed ML model for a given task, nonetheless they can give a baseline model for almost any task in zero time!

We’re living in the future, guys.