Hugging Face adds a decision pipeline: probabilities over allowed answers instead of text
Hugging Face has published a post on what it calls decision models and added support for them to Transformers. The idea is narrow and practical: when an application only needs to pick between answers it already understands, the model should score those answers directly instead of writing out a JSON object token by token.
The interface the post describes has two inputs, a state and a set of named questions, and returns typed decisions with probabilities. Three primitives cover the question types. Choice returns a distribution over named options, such as which team should handle a support ticket. Score is for an ordered scale; it returns the expected value of the level index, so a ticket rated 1.7 on a 0 to 2 urgency scale means the model's probability mass sits mostly on "high", not that it invented a level in between. The third, which the authors call noul after the Bernoulli distribution, returns the probability that a yes/no statement is true and leaves the threshold to the application. Choice and Score also return a confidence value that rescales the winning probability so that a uniform distribution scores 0 and a certain one scores 1.
The post sets this against two familiar tools. A conventional classifier has its labels fixed in the head, so adding a new department means retraining. Zero-shot classification lets labels change per request; decision models go further by taking several typed questions about the same state in one call. Asking an LLM for JSON, the authors argue, can still break the schema, and constrained decoding fixes the structure but not the cost of generating text that is not needed.
On the code side, Transformers now has DecisionPipeline, AutoModelForDecision and DecisionConfig, and transformers serve exposes a /v1/systemone endpoint; the example response reports 148 input tokens and zero output tokens. The same API also works with compatible autoregressive models: Transformers builds the prompts, reads the logits of the answer labels in one batched forward pass and applies softmax, so the interface can be tried without a new model. The aim, the authors write, is a common interface without a common implementation, so that engines such as llama.cpp, the Hub and training tools can treat all such models alike.

Why it matters
Routing, triage and moderation are often done today by prompting a chat model and parsing its output. A standard that returns a probability per allowed answer gives applications something they can threshold and log, and removes a class of parsing failures.