Hugging Face presents decision models for structured AI choices
Hugging Face has published an overview of decision models, an approach intended for applications that need structured choices rather than generated text. Given a state, a set of questions and predefined answers, such a model returns decisions together with probability distributions. The article presents the interface as “State + Questions + Allowed Answers → Decisions + Probabilities.”
The post distinguishes decision models from conventional classifiers and zero-shot classification. Traditional classifiers have fixed output classes, while zero-shot systems can receive candidate labels at inference time. Decision models extend this flexibility to multiple named questions and several answer types. Their primitives include Choice, which selects among alternatives; Score, which evaluates an ordered scale through the distribution’s expected value; and Noul, which estimates the probability of a yes-or-no statement being true.
The article argues that directly selecting among known options can avoid unnecessary autoregressive text generation and some schema-related problems associated with asking an LLM to produce JSON. It demonstrates the approach with the Transformers decision pipeline and the model autotrust/GEV-26B-Decide, using a customer-support example involving billing, urgency and a refund request. Choice and Score responses can also include a confidence measure summarizing how concentrated the predicted distribution is.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.