Hugging Face introduces decision models for structured AI choices

From text generation to direct decisions

Large language models are widely used to answer questions, write code and reason about problems. But Hugging Face argues that some applications do not need a generated explanation or a complete JSON response.

Its example is a customer-support message saying, "I was charged twice for my subscription. I need a refund today." A system processing that message may need to identify the responsible team, determine urgency and establish whether the customer is requesting a refund. An LLM can be prompted to provide those answers in JSON, but a decision model evaluates the available options directly.

Hugging Face describes the core interface as: State + Questions + Allowed Answers → Decisions + Probabilities. The state is the information being evaluated, while questions define what the application wants to know and which answers are available.

What decision models return

In an example using the autotrust/GEV-26B-Decide model through the Transformers pipeline, the system receives a state and a question asking which team should handle a request. The available choices are billing, technical and sales. The result identifies billing as the choice and includes probabilities for all three options, with billing receiving a probability of 0.9982666969299316 in the example.

The interface supports three fundamental primitives. Choice selects one option from several alternatives. Score evaluates something on an ordered scale. Noul estimates the probability that a yes-or-no statement is true.

Hugging Face says a state may be a string, structured JSON or other modalities supported by omni-modal models. Multiple named questions can also be asked about the same state, with each question using its own answer type.

How they differ from existing approaches

Traditional classification models learn a fixed set of labels in a classification head. If an additional department such as Legal is introduced, the existing classifier does not automatically gain that output class, and the model may need changes to its head, additional training data, and retraining or adaptation.

Zero-shot classification allows candidate labels to be supplied at inference time. Hugging Face compares this approach with decision models, which also support runtime-defined answers but extend the format to multiple typed questions, ordered scoring and probabilistic yes-or-no decisions.

The company says decision models do not invent runtime-defined labels or probabilistic classification. Their distinction is a consistent interface that brings these capabilities together, potentially with architectures and training procedures optimized for decisions.

Why the approach matters for applications

Prompting an LLM to produce JSON can result in invalid JSON, unexpected keys or answers outside the allowed choices. Constrained-decoding tools can enforce grammars or JSON schemas, but Hugging Face says they add computational work by restricting the tokens that may be generated.

Decision models target cases where an application already understands the possible answers and needs a reliable structured result rather than an open-ended response. That makes the approach relevant to routing, classification-like workflows and other systems that need probabilities alongside their decisions. Hugging Face published the explanation on November 10, 2026.