Towards Foundation Decision Models: Drex
Nischay Dhankhar, Alisher Amantay, Zhanibek Datbayev, Darkhan Akhmetov, Yash Shah, Nurmukhamed Ubaidullayev, Raman Saparkhan, Dos Baha and 1 others
Nace.AI
Published November 5, 2026 · on the feed November 6, 2026 · ▲ 24 · arXiv 2611.07854
The abstract
Autoregressive approaches with large language models, such as long chain-of-thought reasoning, have become the default for new AI applications because they reach higher accuracy on a wide range of tasks. In production, however, they give inconsistent answers, require parsing of outputs, and add cost and latency. Decision models address this: given a question and a set of options supplied by the user, they return a probability distribution over the options without autoregressive decoding. We present Drex, a family of four open decision models with 8B (DLM), 9B (v1.5), 35B with 3B active (v2.0), and 27B (Max) parameters. Each request consists of a state (a document, record, or agent context, optionally with images) and questions about it, each asking for a choice among options, a yes/no judgment, or a score on an ordered scale. Training combines supervised fine-tuning with reinforcement learning from adjusted feedback (RLAF), plus vision tuning for Max. Ablations show that RLAF improves accuracy and that consistency losses improve robustness to option order. On Decision Index 0.4.1, Max sets a new state of the art among the decision models we compare (67.52 on the public suite vs. 60.70 for Jev 1.14.0). On Decision Bench v4, it is statistically tied with d3-turbo and GPT-6 Sol and trails Claude Opus 5.5 by 1.6 points, and it has the highest accuracy among decision models on DecideBench. Max answers a typical request in 183 ms, against about 2.0 s for GPT-6 Sol and 3.5 s for Claude Opus 5.5, and its outputs are stable under option reordering. All four models are open source: https://github.com/nace-ai/drex-decision-models
The Brief
2:41A short podcast summary of the paper. Written and voiced by AI.
Transcript
Towards Foundation Decision Models. Drex. By Nischay Dhankhar and eight others, from Nace AI. Published on the fifth of November.
Large language models that work autoregressively, long chain-of-thought reasoning among them, have become the default for new AI applications, because they reach higher accuracy across a wide range of tasks. Production is less kind to them. They give inconsistent answers, their outputs have to be parsed, and they add cost and latency. Decision models are a response to that. Give one a question and a set of options supplied by the user, and it returns a probability distribution over those options directly, with no autoregressive decoding at all.
Drex is a family of four open decision models of this kind, spread across four different sizes, each answering in that same direct way.
The four members are these. DLM has eight billion parameters. Version one point five has nine billion. Version two point zero has thirty-five billion parameters, of which three billion are active. And the one called Max has twenty-seven billion.
Every request has the same structure. There is a state, which can be a document, a record, or an agent's context, optionally with images. Then come questions about that state. Each question asks for one of three things: a choice among options, a yes or no judgment, or a score on an ordered scale. The output is a distribution, not a paragraph.
Training combines supervised fine-tuning with reinforcement learning from adjusted feedback, which the authors abbreviate RLAF. Max additionally receives vision tuning. Ablations back two of the design choices: RLAF improves accuracy, and consistency losses make the models more robust to the order in which options are presented.
These models are aimed at the production setting the authors describe, where inconsistent answers, parsing, cost and latency are the pain. On Decision Index, Max sets a new state of the art among the decision models compared, scoring sixty-seven point five two on the public suite against sixty point seven for Jev.
On Decision Bench version four, Max is statistically tied with d3-turbo and with GPT-6 Sol, and trails Claude Opus five point five by one point six points. On DecideBench it has the highest accuracy among decision models.
Latency is where the format pays off. Max answers a typical request in one hundred and eighty-three milliseconds. The comparison figures are about two seconds for GPT-6 Sol and three point five seconds for Claude Opus five point five. Its outputs also stay stable when the options are reordered.
All four models are released openly. Drex rests on a reframing: treat a production question as a choice among supplied options, answer it with a distribution instead of generated text, and the reported results put Max level with or close behind the general models it was compared against on accuracy, while answering in a fraction of their time.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Reasoning models are accurate, but production asks for more than accuracy.
You run a product where a language model answers questions like 'approve or reject?' or 'score this from one to five' over documents and records.
The road ahead
- 1 · The problem ~6 min Production needs clean, consistent, cheap answers to questions whose options change every time, and long chain-of-thought text does not give them.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Drex reads a state and questions with user-supplied options, and returns a probability for each option instead of writing text.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Drex is trained in stages: supervised fine-tuning, then reinforcement learning from adjusted feedback, and for Max a final vision tuning stage.
- 6 · The experiment ~6 min You’ll guess: On Decision Bench v4, how did Max, the largest open model in the family, compare with Claude Opus 5.5?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.