Decision models evaluate a given state against a set of predefined answers and return structured decisions with probability distributions, rather than producing open‑ended text. A November 10, 2026 post describes this approach and contrasts it with prompting LLMs to emit JSON or using conventional classifiers.
How decision models work
The core interface is: State + Questions + Allowed Answers → Decisions + Probabilities. Three fundamental primitives are described: Choice (a categorical distribution over options), Score (an expectation over an ordered scale) and Noul (a Bernoulli probability for yes/no questions).
Examples in the post use a customer message such as “I was charged twice for my subscription. I need a refund today.” For a Choice question routing to teams, the model can return probabilities like billing 0.998, technical 0.001, sales 0.001 and report the highest probability as the choice. For an ordered urgency scale with levels 0=low, 1=medium, 2=high, the model can output probabilities 0.01, 0.26 and 0.72, producing an expected score of 1.7. For a Noul question such as whether the customer requests a refund, the model may return a probability like 0.9947.
The post also describes a normalized confidence measure that summarizes how concentrated the model’s belief is: a uniform distribution maps to confidence 0 and a distribution entirely on one option maps to confidence 1. In the billing example, a winning probability of 0.998 yielded a reported confidence of 0.997.
Interface and ecosystem support
The post notes that decision models sit between fixed-label classifiers and zero‑shot classification: they accept labels at inference time and expose typed questions and probabilistic outputs as a consistent contract. The intent is to let different implementations coexist while exposing the same runtime API so tooling can be shared.
Concrete tooling updates described include support in transformers for DecisionPipeline, AutoModelForDecision and DecisionConfig. The examples reference models such as autotrust/GEV-26B-Decide and show how DecisionConfig.from_pretrained and AutoModelForDecision can be used to inspect and load models. The post also mentions that inference engines like llama.cpp can serve decision models through a common API and that the Hub can package them consistently.
Original source: Hugging Face Blog