NTH

Towards Foundation Decision Models: Drex

AuthorsNischay Dhankhar, Alisher Amantay, Zhanibek Datbayev, Darkhan Akhmetov, Yash Shah, Nurmukhamed Ubaidullayev, Raman Saparkhan, Dos Baha, Abulhair Saparov

AffiliationsDarkhan Akhmetov · Raman Saparkhan

November 8, 2026 3 min read
Watch on YouTube
The one-line take

Drex aims to make AI decisions faster and more consistent by predicting directly among options instead of generating long answers.

Key results

67.52
Decision Index 0.4.1 public score

Drex Max's chance-corrected score; Jev 1.14.0 scores 60.70.

183 ms
Drex Max request latency

Typical single-request latency, versus about 2.0 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5.

98.3%
Image-only Decision Bench v4 accuracy

With image input; accuracy is 63.3% when Max receives text renderings instead.

2.0
RLAF Decision Index gain

Point increase over SFT alone in the controlled Drex Max ablation.

4.2%
Option-order flip rate after consistency losses

Reduced from 6.3% without the consistency losses.

1.15M
Text training records

Verified decision records in the shared text training mix.

What the paper found

Drex is an open-weight family of decision models that returns calibrated probabilities over options supplied at inference time, without generating text or parsing answers. Its pointer models encode a state once and reuse it across questions; Drex Max instead uses a 27B Qwen3.8-27B backbone, bidirectional full-attention layers, and a readout head, trading shared caching for stronger decisions and image input. Training combines supervised fine-tuning with reinforcement learning from adjusted feedback, or RLAF, and a vision-tuning stage; the shared text training mix contains 1.15M verified records. On Decision Index 0.4.1, Drex Max scores 67.52 on the public suite, ahead of Jev 1.14.0 at 60.70. On Decision Bench v4, its image encoder raises accuracy on image-only questions from 63.3% using text renderings to 98.3%. In a controlled ablation, RLAF adds 2.0 points to the Decision Index score, while consistency losses reduce the option-reordering flip rate from 6.3% to 4.2%. A typical Drex Max request takes 183 ms, compared with about 2.0 seconds for OpenAI’s GPT-6 Sol and 3.5 seconds for Anthropic’s Claude Opus 5.5. The results position Drex as a fast alternative for structured choices, though its performance still trails frontier chat models on some knowledge-heavy tests.

Original 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

Read the original paper

More in Foundation Models

Browse all 53 papers →
02Foundation Model

Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage

Chenfan Qu, Yachao Liang, Mingqi Fang, Jiangwei Xie, Zhenming Wang, Baoyue Zhang, Changjiang Jiang, Zheng Chang, Fanli Jin, Xuekang Zhu, Kaiwen Feng, Junchi Li, Weihong Zhang, Haoyue Wang, Chenfeng Zhang, Yinan Xia, Fangming Cui, Dong Fan, Jingjing Liu, Tianyi Zhang, Luobin Chen, Ji-Zhe Zhou, Sheng Li, Wei Wang, Bin Li, Ran He, Xinpeng Zhang, Linna Zhou, Jian Liu, Weiqiang Wang

Venus-DefakerOne2 aims to detect many kinds of image and video forgery with one broadly trained vision-forensics model.

Read analysis
03Foundation Model

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI

NVIDIA, :, Aastha Jhunjhunwala, Aayushi Ashokkumar Srivastava, Abhibha Gupta, Abhinav Garg, Abhinav Khattar, Achsah Wildish, Adam Rajfer, Adi Renduchintala, Adil Asif, Adriana Flores Miranda, Aileen Zaman, Aishwarya Chhabra, Akanksha Shukla, Akhiad Bercovich, Alejandra Rico, Aleksa Gordic, Aleksander Ficek, Alex Gronskiy, Alex Steiner, Alexander Bukharin, Alexandre Milesi, Ali Roshan Ghias, Ali Shehper, Ali Taghibakhshi, Alice Gatti, Alisa Liu, Alok Kumar, Alok Tripathy, Amala Sanjay Deshmukh, Aman Miryala, Aman Shanbhag, Ameya Sunil Mahabaleshwarkar, Ameya Velingker, Amir Klein, Amir Samani, Amit Zuker, Amnon Geifman, Anahita Bhiwandiwalla, Ananth Subramaniam, Andre Manoel, Andrea Santilli, Andrew Tao, Andrii Skliar, Anil Thomas, Anish Mahishi, Anna Shors, Anna Warno, Anuj Doshi, Anusha Kamath, Anwith Kiran, Ao Tang, Arham Mehta, Arnab Dhara, Arnav Komaragiri, Arnav Kundu, Arti Jain, Arun Rao, Arun Venkatesan, Arushi Goel, Asha Anoosheh, Ashish Vaswani, Ashwath Aithal, Asif Ahame...

Nemotron 3.5 pairs hybrid Mamba-Transformer models with agent training and inference optimizations to deliver capable, efficient open AI systems.

Read analysis