Hugging Face adds Decision Models to Transformers library

Hugging Face has added a new class of models to its Transformers library aimed at structured decision tasks where an LLM's free-form generation is overkill. The November 10 announcement introduces 'decision models,' which take a state plus a set of predefined questions and return structured decisions with probabilities rather than generating text token by token.

The framing is a direct swipe at routing and classification by LLM. 'LLMs are not guaranteed to follow the requested schema,' the post notes, and constrained-decoding libraries such as Outlines and XGrammar, while helpful, 'introduces additional work.' Decision models skip the autoregressive loop entirely.

The framework exposes three primitives. Choice is described as 'the closest to traditional classification.' Score 'computes the expected value of the level index' on an ordered scale. Noul returns a Bernoulli probability for a yes/no question; its name, the authors write, 'comes from Bernoulli.' New Transformers classes including DecisionPipeline, AutoModelForDecision and DecisionConfig ship with a `/v1/systemone` serving endpoint, and llama.cpp can serve GGUF decision models through the same API.

The post publishes no benchmarks against LLMs or fine-tuned classifiers. On intent, the authors are explicit: 'Our goal at Hugging Face is to agree on a common interface without forcing a common implementation.' Fine-tuning works today through the Transformers Trainer API; TRL support is listed as incoming. Two researchers on our Who's Who radar had shared the link by the time the post went up.