Hugging Face adds a decision pipeline for models that score options instead of writing text
Transformers gains a "decision" task that returns a probability over predefined answers, as a new arXiv paper asks how such models fit into long-horizon agents.
Hugging Face has published a post, dated 10 November 2026, describing a new decision task in its Transformers pipeline API. The authors (Aritra Roy Gosthipaty, Lysandre, Pedro Cuenca and Tom Aarsen) frame it around a familiar pain point. Teams often prompt a large language model for a JSON answer to a routing or triage question, then parse the text. A decision model skips the generation step. It takes a state (such as a support ticket), one or more questions, and a fixed set of allowed answers, and returns a chosen answer with a probability for each option.
The worked example routes a double-charge complaint to one of three departments. The post shows the model, autotrust/GEV-26B-Decide, assigning the billing option a probability above 0.99. The post does not state that model’s parameter count, context length or licence.
The post places the idea between two older approaches. A conventional classifier has its labels fixed in its output head, so adding a fourth department means retraining. Zero-shot classification with a natural language inference model, such as facebook/bart-large-mnli, accepts new labels at inference time by treating each label as a hypothesis. Decision models are presented as a step beyond that, with a structured interface of state, questions and criteria.
The naming is settling. TypeSafe AI released a model called Jev and described it as a “System One model”, after the fast-versus-slow thinking distinction. Hugging Face says the community has instead converged on “decision model”. Separately, a paper by Haochen Liu and nine co-authors, submitted to arXiv on 12 November 2026, asks whether such models serve agents best as actor, judge or teacher.
Why it matters
For practitioners, this is structured output without the parsing. Calibrated probabilities over known options are easier to threshold, log and monitor than free text, and routing becomes a single forward pass. The interface also keeps label sets flexible without retraining a head.
Be sceptical on two fronts. The post’s example scores are the vendor’s demonstration on one ticket, not a benchmark, and no independent evaluation is cited. Model details for the example checkpoint are not given. Watch whether the task gains more open checkpoints, how confidence values hold up on out-of-distribution inputs, and whether the arXiv work on agent roles produces measurable gains over prompting an LLM for JSON.