DEV Community

aimodels-fyi
aimodels-fyi

Posted on Originally published at aimodels.fyi

Drex Maps Changing Option Sets to Probabilities in a Single Model Pass

This is a Plain English Papers summary of a research paper called Drex Maps Changing Option Sets to Probabilities in a Single Model Pass. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.

Drex turns option sets into direct probability outputs

Drex Max is a 27B-parameter decision model that returns probabilities over user-supplied options without autoregressive decoding. A request pairs a state—such as a document, record, or agent context, optionally with images—with choice, yes/no, or ordered-score questions. The result is a distribution over the allowed answers, rather than generated text that must be parsed.

The family contains four models: DLM at 8B parameters, v1.5 at 9B, v2.0 at 35B with 3B active, and Max at 27B. All share an interface that supports up to 255 options per question and can process multiple questions about one state. The paper targets production concerns it names directly: inconsistent generated answers, output parsing, latency, and inference cost. A concrete application of the same direct-option interface is entity matching, where yes/no probabilities can replace generated labels.

Drex is aimed at decisions whose candidate set changes at inference time, unlike a classifier with a fixed label head. Its probabilities also support confidence-based selection and routing, while score questions return an expected level alongside probabilities for each level. The paper evaluates accuracy, calibration, option-order stability, and image grounding rather than treating a valid output as sufficient.

How Drex differs from generation and earlier decision models...


Continue reading the full paper summary on AIModels.fyi →

Top comments (0)