Hugging Face introduces shared decision-model interface with probabilities instead of generated text
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Agents no longer need to compose text for classification and routing: an application can receive a typed decision and its probability directly, faster and in a consistent format across different models.
Hugging Face has standardized support for decision models, which evaluate permitted decisions directly rather than writing a response token by token. A request consists of context and typed questions: choice selects an option and returns a probability distribution, score calculates an expected value on an ordered scale, and noul gives the probability of a yes-or-no answer. This allows one call to route a request, assess urgency and check a condition without parsing potentially malformed JSON or paying to generate an explanation; the API example returns zero output tokens. Transformers has gained DecisionPipeline, AutoModelForDecision and DecisionConfig, while Transformers and llama.cpp support the shared /v1/systemone endpoint. The compatible interface works with purpose-trained models and with ordinary autoregressive LLMs, whose option logits are normalized without text generation. The Hub has added a dedicated tag for finding such models, and the Trainer API already supports fine-tuning them.
Sources 1