Hugging Face introduces decision models for structured, probability-based outputs
New interface enables applications to query models for structured decisions and confidence scores without text generation.
1 source · cross-referenced
- Hugging Face describes a new class of models called 'decision models' that return structured decisions and probabilities instead of generating text.
- The interface accepts a state, a set of questions with predefined primitives, and returns decisions with confidence scores.
- Decision models aim to avoid common pitfalls of using LLMs for structured tasks, such as schema drift and unnecessary autoregressive overhead.
Hugging Face introduced a new interface for models it calls 'decision models,' which take a state and a set of questions with predefined answer primitives, then return structured decisions and associated probabilities rather than generating free-form text.
The company argues that many applications do not require open-ended text generation but instead need reliable, structured outputs—such as routing a customer support ticket to a specific team, assessing urgency, or determining if a refund is requested.
In the example provided, a decision model evaluates a support ticket and returns a JSON-like structure with a department choice ('billing'), a confidence score (0.9974), and a probability distribution over possible departments, without generating any text.
The post contrasts this approach with traditional classification and zero-shot classification, noting that decision models allow the set of possible answers to change at inference time without retraining, and can handle multiple named questions about the same input state.
Hugging Face also outlines how to serve decision models using Transformers and llama.cpp, discover them on the Hugging Face Hub, and train them using Hugging Face tooling, positioning them as a new tooling category within its ecosystem.
The interface is designed to avoid common issues when repurposing LLMs for structured tasks, such as schema drift (e.g., invalid JSON or unexpected keys) and the computational overhead of constrained decoding or schema-enforcing libraries like Outlines or XGrammar.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79