Hugging Face adds decision-model support to Transformers

Hugging Face has added decision-model support to Transformers, giving developers a common interface for asking compatible models for typed choices, ordered scores and yes-or-no probabilities without generating response text. The release adds DecisionPipeline, AutoModelForDecision and DecisionConfig, together with a serving route and a Hub tag for finding compatible models.

A request combines a shared state with one or more named questions. The interface returns structured answers and probability distributions over the allowed responses. For compatible autoregressive language models, Transformers builds the decision prompts, evaluates answer-label logits in batched forward passes and applies softmax, rather than decoding an open-ended answer token by token.

Each of the three question primitives serves a different need. A choice question selects among named alternatives and returns the probability for each option plus a confidence value. A score question represents an ordered scale and reports the probability-weighted expected value of the level index, alongside the distribution, legend and confidence. A noul question handles a binary statement by returning the probability that it is true; the application decides what action, if any, to take from that probability.

Developers can invoke the feature through pipeline("decision", model=...) in Transformers, or load a model and its configuration with AutoModelForDecision and DecisionConfig. Transformers Serve exposes the same request format at POST /v1/systemone. The endpoint wraps the served model in DecisionPipeline, so the pipeline and HTTP route build the same prompts and return the same answer types. According to the serving documentation, it supports autoregressive models that have a chat template.

Hugging Face’s published support-ticket example asks which department should handle a case, rates its urgency and estimates whether a refund is being requested. The response contains all three typed answers and their probabilities while reporting zero output tokens, reflecting that the model scored allowed answers instead of generating response text.

The interface is not tied to a single architecture. Hugging Face says it can cover models with dedicated decision heads as well as systems built from generative-model backbones. A dedicated decision-model filter on the Hugging Face Hub gives developers a way to find compatible models.

The Transformers pipeline documentation lists decision as a supported pipeline task but labels decision models, DecisionPipeline and DecisionConfig experimental. It says those APIs may change at any time without deprecation.