Drex Turns Structured AI Decisions Into a Single Forward Pass

Introduction

Large language models are often used for moderation, routing, ranking, and agent control. Yet many of these applications do not need a paragraph of generated text. They need a choice from a known set of options, together with confidence values that software can consume directly. Generating an answer token by token adds latency, creates parsing requirements, and can make otherwise identical requests behave inconsistently.

Nace.AI’s Drex approaches this problem as a dedicated decision-model task. Rather than producing an autoregressive response, it returns a probability distribution over options in a single forward pass.

A structured decision interface

Each Drex request contains a state and one or more questions about that state. The state may be a document, record, or agent context, and it can optionally include images. A question can ask the model to select among user-supplied options, make a yes-or-no judgment, or assign a score on an ordered scale.

This interface is designed for production pipelines. Applications can consume probabilities directly, set confidence thresholds, trigger human review, or route uncertain cases to a larger model. They do not have to depend on regular expressions or fragile conventions for extracting an answer from free-form text.

Four models and a different training recipe

The Drex family includes four variants: the 8B DLM, 9B v1.5, 35B v2.0 with approximately 3B active parameters, and the 27B Max model. Training combines supervised fine-tuning with reinforcement learning from adjusted feedback, or RLAF. The method blends verified outcomes with signals from a reasoning judge while applying a KL constraint. Max also receives vision tuning for image-containing inputs.

The reported ablations point to two practical effects. RLAF improves accuracy, while consistency losses reduce sensitivity to the order in which options are presented. In the cited experiment, option-order flips fall from 6.3% to 4.2%. That matters for a decision API: rearranging a list should not change the model’s view of the underlying question.

Accuracy and latency

On the public Decision Index 0.4.1 suite, Drex Max scores 67.52, compared with 60.70 for Jev 1.14.0. It retains a 66.60 score on maintainer-run private tests. On Decision Bench v4, Max is statistically tied with d3-turbo and GPT-6 Sol, while trailing Claude Opus 5.5 by about 1.6 points. The paper also reports the highest accuracy among decision models on DecideBench.

The efficiency difference is central to the proposal. A typical Max request takes about 183 milliseconds, versus roughly 2.0 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5 in the cited comparison. Actual production performance will still depend on hardware, batching, input length, and serving configuration.

Why it matters

Drex suggests that not every AI workflow should be built around a general-purpose chat model. When the task has a defined state, a known decision space, and a need for structured output, a specialized probability-producing model may be easier to integrate and cheaper to operate. Potential uses include document review, risk triage, search-result ranking, tool selection for agents, and multimodal record assessment.

There are also clear boundaries. A decision model does not replace open-ended reasoning, and a probability over supplied options is not automatically a calibrated estimate of real-world risk. Results depend on how the state, question, and options are designed. Teams will need to evaluate calibration, business-specific error costs, and behavior under distribution shifts rather than relying only on public benchmarks.

Drex is released with open weights and inference code, giving developers an opportunity to test whether this narrower interface can provide a better balance of speed, consistency, and accuracy for their own systems.

Source: Hugging Face Daily Papers