Open decision models that return only probabilities over options ── A preprint proposing Drex, which scores user-supplied options without generating text, and reporting its accuracy and latency

1. What problem does the paper address?

The problem is that even routine judgments inside business workflows are being handed to text generation. Having a large language model write a long chain of thought before answering has become the default for new AI applications, because it reaches higher accuracy on a wide range of tasks. The authors point out three problems that show up in production^[1]: answers are inconsistent across repeated requests, the output has to be parsed to pull out the answer, and cost and latency go up.

Many judgments that businesses ask of AI are not free-form writing. Is anything missing from this document? Which category does this record fall into? Where on a graded scale does the risk in this wording sit? The candidate answers are fixed in advance. A generative model still writes prose, and someone still has to read a yes or a no out of that prose. The paper discussed here is a preprint posted on arXiv and has not been peer reviewed, so what follows should be read as the authors' claims.

2. What does it propose?

The core idea is a decision model that returns a probability distribution over user-supplied options directly, without generating text token by token^[1]. The authors call it Drex and release it as a family of four open models of different sizes: DLM at 8B parameters, v1.5 at 9B, v2.0 at 35B total with 3B active, and Max at 27B.

Each request pairs a state with questions. The state is a document, a record or an agent's working context, optionally with images. Each question asks for a choice among options, a yes/no judgment, or a score on an ordered scale^[1]. According to the repository, every option gets a probability back and nothing is generated^[2].

Training combines supervised fine-tuning with a form of reinforcement learning the authors call RLAF (reinforcement learning from adjusted feedback), plus vision tuning for Max. Consistency losses are added so that shuffling the order of the options does not change the answer^[1].

3. What does it show, within the abstract only?

The abstract reports the following^[1].

• Training ingredients. In ablations, RLAF improved accuracy and the consistency losses improved stability under changes in option order.
• Against other decision models. On the public suite of Decision Index 0.4.1, Max scored 67.52 against 60.70 for Jev 1.14.0, which the authors describe as a new state of the art among the decision models they compare. It also had the highest accuracy among decision models on DecideBench.
• Against large generative models. On Decision Bench v4, Max was statistically tied with d3-turbo and GPT-6 Sol and trailed Claude Opus 5.5 by 1.6 points.
• Latency. Max answered a typical request in 183 ms, against about 2.0 s for GPT-6 Sol and about 3.5 s for Claude Opus 5.5, and its outputs stayed stable under option reordering.

One point deserves attention: the abstract does not say Max beats the strongest generative models on accuracy. The claim is closer to "comparable or slightly lower accuracy, much faster, and unaffected by option order".

4. A concrete case

Consider a pharmaceutical company that wants AI to help with a first-pass check of promotional materials before they reach the review department. A reviewer and a systems lead discuss it.

What changes in this example is the shape of the work: people fix the question and the options in advance. Writing down what counts as "yes" in the wording of the options is itself a way of putting the review criteria on paper.

5. What is not new, and where are the limits?

The format of supplying options and getting probabilities back is not new. The abstract itself compares against Jev 1.14.0, an existing decision model of the same kind. What the authors present as new is the release of four open models at different scales, the combination of RLAF with consistency losses, and a Max model that reads images.

Even within the abstract, several limits are visible.

• No answers outside the options. Users supply the options. If the right answer is not among them, the model cannot point to it. An earlier preprint lists distinguishing "no answer fits" from uncertainty among candidates as a requirement for decision models^[4]; the Drex abstract does not say how this is handled.
• The benchmarks are opaque. The abstract does not describe the tasks in Decision Index, Decision Bench or DecideBench, or who built them. "Best among the decision models we compare" also means the authors chose the comparison set.
• "Statistically tied" is undefined. Which test was used and what margin counts as equivalent is not stated in the abstract.
• Latency conditions are missing. The length of a "typical request" and the hardware used are not given. The generative models' times also depend on how their providers operate them.
• Ordered scales. An earlier preprint reported that decision models of this kind underuse graded rating scales, concentrating their judgments on fewer labels than the gold distribution^[5]. The Drex abstract does not address this.
• What "open source" means here. The repository lists three models under the Nace.AI Open RAIL-M license and DLM under CC BY-NC^[2]. These can restrict commercial or certain uses, so the license terms need reading before any business use.

6. Why is this topic clustering now?

Strictly speaking, it is not clustering yet. In the collection system the cluster size is 1, and no related papers were found. The attention signals are 15 upvotes on Hugging Face and 98 GitHub stars, so the only signal that fired comes from implementers. These numbers show interest, not that the paper's claims are correct.

The decision-model format has, however, been showing up repeatedly on arXiv over recent weeks. One preprint proposed a decision-only design whose answer does not change under reordering and which can signal that no candidate fits^[4], and another reported bias in how such models use graded scales^[5]. Within that stream, Drex puts its weight on releasing runnable weights at several scales. The repository and model page publish the steps for running the models locally and where to get the weights^[2][3].

The keywords center on avoiding autoregressive decoding, probability distributions over options, and stability under option reordering. Read together, they describe an effort to split off a component that returns fixed-format judgments quickly and consistently, separate from the push for accuracy with generative models.

7. Where does this connect to pharma and regulation?

There are three points of contact. The first is reproducibility. A system that gives a different answer each time the same material, question and options are submitted is hard to defend as a review record. A decision model whose answer does not depend on option order, and whose output is a stored probability, is easier to document. Stable output is still not the same as a correct judgment.

The second is responsibility for the questions and options. A decision model uses the wording of the options as its criteria. How to define "wording that suggests an unapproved indication", or which graded scale to use, is decided by the review department, not the model. If the definition is vague, fast and stable judgments will pile up on a vague basis.

The third is volume and speed. If Max's latency holds as reported, one use would be running a first-pass check over large numbers of materials or records and narrowing down what people need to look at. Because Max reads images, the check could include figures and tables in materials. In every case, the model's judgment comes before human review rather than replacing it. Until accuracy is re-measured on in-house materials and the license terms are confirmed, keeping it at the evaluation stage is the sound course.