Towards Foundation Decision Models: Drex

AI summary

Overview

Research area: Natural Language Processing — specifically decision models, a class of models that return a probability distribution over a user-supplied set of options without autoregressive decoding.

Technical level: Advanced. The paper is a technical report with architecture diagrams, loss derivations, and paired statistical evaluation; the framing and results are readable by a motivated non-specialist, but the training and inference details assume familiarity with transformer architectures and reinforcement learning.

Scope (one sentence): The paper introduces Drex, a family of four open decision models (8B DLM, 9B v1.5, 35B-with-3B-active v2.0, and 27B Max), and studies their calibration, sensitivity to option order, training pipeline, and architecture choices against frontier chat models and other decision models.

What This Paper Is About

Autoregressive LLMs with long chain-of-thought reach high accuracy, but in production they give inconsistent answers, require parsing of generated text, and add cost and latency. Decision models address this by taking a question plus a set of options supplied by the user and returning a probability over those options directly, with no decoding. The paper asks whether a single model can serve arbitrary option sets with calibrated probabilities, answer many questions over one long context, stay stable when options are reordered, and still run at production latency — and reports four Drex models built to test that question.

Key Contributions

1. Calibration at scale (RQ1). The paper measures the calibration of Drex Max, v2.0, v1.5, and a commercial baseline on about 230k decisions each, reporting bootstrap confidence intervals, breakdowns by domain and by number of options, selective accuracy, and the effect of a single temperature fitted on held-out data.

2. Order-sensitivity and confidence analysis (RQ2). Four production models, accessed through their public APIs, are tested on 1,591 decisions each presented in three orders (original, reversed, seeded random permutation). The paper finds that the top option changes with order mostly on decisions with many options and low model confidence, and that declining to answer order-unstable decisions is no more accurate than declining the lowest-confidence ones.

3. A three-stage training recipe (RQ3). Supervised fine-tuning (SFT) with two consistency terms, then reinforcement learning from adjusted feedback (RLAF) with a KL penalty relative to the SFT checkpoint, then SFT of the image encoder and its projection into the language model. An ablation on Drex Max's own backbone and data shows RLAF adds 2.0 points on the Decision Index, the consistency terms lower the flip rate from 6.3% to 4.2%, and the vision stage does not reduce accuracy on text inputs.

4. Drex Max and one interface across four backbones (RQ4). Max is a dense multimodal model trained from Qwen3.8-27B that removes the causal mask in its full-attention layers and answers with a readout head (a bias-free linear layer mapping the final hidden state to 255 logits, one per option code). The same request/response interface is implemented on four different backbones: a Gated DeltaNet hybrid, a diffusion language model with bidirectional attention, a sparse mixture-of-experts model, and the dense multimodal 27B model.

The paper also contributes a paired evaluation protocol across Decision Index (0.2.1, 0.4.1, and 0.4.1 with private tests), Decision Bench v4, JevBench, DecideBench, a game arena, a computer-use judging probe, and a vision suite, with Wilson intervals, paired bootstrap intervals, and exact McNemar tests.

Main Findings

• State-of-the-art on Decision Index 0.4.1 among compared decision models: Drex Max scores 67.52 on the public suite versus 60.70 for Jev 1.14.0. Drex DLM and GPT-6 Luna Decisions have no full score (the full score includes private tests).

• Second place on the independent Artificial Analysis Decision Index: Max scores 64, against 65 for the three leading models.

• Mixed but competitive on Decision Bench v4: Max beats Jev 1.14 and the other Drex models, is statistically tied with d3-turbo and GPT-6 Sol, and trails Claude Opus 5.5 by 1.6 points.

• Best decision-model accuracy on DecideBench and competitive on JevBench-231: Max has the highest accuracy among decision models on DecideBench and is competitive with the top model on JevBench-231.

• Large latency advantage: 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. At launch, median latency was 183 ms for Max, 151 ms for v2.0, and 145 ms for v1.5.

• Order sensitivity is concentrated in hard, many-option, low-confidence cases: The top option changes with option order mostly on decisions with many options and low confidence. Averaging the three predicted distributions across the three orders improves log loss but not accuracy.

• Consistency losses reduce flip rate: Adding the permutation-consistency and state-shuffle-consistency terms lowers the flip rate from 6.3% to 4.2% in the Max ablation.

• RLAF adds accuracy: The ablation shows a 2.0-point Decision Index gain from the RLAF stage.

• Vision training is not a text-accuracy tax: The third training stage does not reduce accuracy on text inputs; the paper separately measures how accuracy changes when the model receives an image itself rather than a text description and transcription.

• Pointer models are cheaper per request; Max is more accurate per question: Because Max's full-attention layers are non-causal, the state representation depends on the question and cannot be cached across questions, so Max pays for the state once per question while the pointer models pay once per request. For many questions over long documents, the pointer models remain cheaper.

• Architectural guarantees vs. learned properties: Closed-world output (probabilities supported only on supplied options) and question independence are guaranteed by construction; order robustness, calibration, and grounding in every supplied modality are learned and therefore measured rather than assumed.

• Caveat on causality: Outside the one ablation of Max's training, the paper reports associations and trade-offs rather than causal relationships, and Section 10 lists comparisons not run and hypotheses needing further experimentation (including which specific design and training changes drive Max's gains).

Methodology in Plain English

Each request is a state (a document, record, or agent context, optionally with images) plus one or more questions. A question is one of three types: a choice among up to 255 named options, a no/yes judgment with user-supplied definitions, or a score on an ordered integer scale where the model also returns the expected level. The model returns a probability for every option and nothing is generated token by token.

The three pointer models (DLM, v1.5, v2.0) pack the state and all question blocks into one token sequence under a block-causal mask: each question sees the whole state but no other question, so the state is encoded once and cached. A small pointer head turns the hidden state at a <decide> token into a query and the hidden state at each option's closing tag into a key, then takes one softmax over exactly the supplied options. This head is tiny — 2.1M parameters for v1.5 (hidden size 4096) and 1.0M for v2.0 (hidden size 2048), under 0.03% of either backbone — and the temperature is fixed at 1.

Drex Max takes the other trade: each question is its own sequence containing images, state, instruction, options, and a final <ans> marker, with the causal mask removed in the 16 full-attention layers of its 64-layer Qwen3.8-27B backbone (every fourth layer). Options are tagged with single-token codes (A, B, … up to 255), and a bias-free linear layer maps the <ans> hidden state to 255 logits; only the logits of the K supplied options are kept and normalized, with temperature fitted post-training at 1.009.

Training has three stages. First, full-parameter supervised fine-tuning on 1,150,460 verified text decision records across eleven domains, with a cross-entropy term plus an ordinal (squared Cramér distance) term for score questions, and two consistency penalties — one for permuting options, one for shuffling state sentences or fields. Second, RLAF, which replaces one-hot targets with an adjusted target built from a teacher distribution and a verified outcome (gold answer, execution result, or engine output) and maximizes a proper scoring rule against that target, regularized by KL divergence from the SFT checkpoint. Third, for Max only, SFT of the image encoder and its projection into the language model.

Evaluation is deliberately paired: the same samples are run through Drex models and frontier chat models, with Wilson intervals for accuracies, paired bootstrap intervals for differences, and exact McNemar tests.

Why This Matters

Impact on research. The paper separates the properties a decision model should have — closed-world output, question independence, order robustness, calibration, and grounding in every supplied modality — and shows which are guaranteed by architecture versus which must be learned and measured. It also provides one of the more careful public treatments of option-order sensitivity and calibration on real decision workloads, and an ablation isolating the contribution of each training stage.

Real-world applications (as framed by the paper's interface and evaluations):

• High-volume classification and routing where the set of labels varies per request, so a fixed classification head cannot be used.
• Agent and document workflows that ask many questions over one long context (up to 128k tokens), where caching the encoded state cuts cost.
• Vision-grounded decisions, where the model must read the image rather than a text transcription of it.
• Judging and scoring tasks on integer scales, such as the paper's computer-use judging probe and game arena, where an expected value is more useful than a full distribution.
• Confidence-thresholded routing, since calibration makes a probability cutoff usable as a routing rule.

Industry relevance. The headline practical claim is cost and latency: matching frontier models on well-defined tasks at a fraction of the cost, 183 ms versus roughly 2.0 s and 3.5 s. Removing output parsing and guaranteeing a valid answer from a closed option set matters for production systems. All four models are open weight — Max, v2.0, and v1.5 under the Nace.AI Open RAIL-M license, DLM under the CC BY-NC 4.0 license of its backbone, and repository code under Apache-2.0 — with safetensors, GGUF builds, llama.cpp and Ollama runtimes, plus hosted API and OpenRouter access.

Future Directions

• Which design changes actually drive Max's gains? The paper explicitly flags this as an open hypothesis requiring further experimentation, since most evidence is comparative rather than causal.
• Closing the remaining gap to the top model. Max trails Claude Opus 5.5 by 1.6 points on Decision Bench v4 and sits 1 point behind the three leading models on the Artificial Analysis Decision Index; whether these gaps come from capacity, training data, or the readout formulation is unresolved.
• Order robustness as a learned rather than guaranteed property. Max's readout head must bind option codes to option text, making order robustness entirely learned; better ways to secure it — or to exploit the log-loss gain from averaging over permutations without losing accuracy — remain open.
• Extending modality coverage. Video is handled only by sampling frames and passing them as images, since the model has no temporal encoder.
• The unrun comparisons and open hypotheses listed in Section 10, which the authors say are needed before stronger causal claims can be made.

Target Audience

Engineers and researchers building production classification, scoring, or routing systems on top of language models; teams weighing autoregressive prompting against discriminative inference on cost and latency grounds; and anyone studying calibration, option-order sensitivity, or reinforcement-learning post-training for structured outputs. Readers wanting only the practical bottom line can read the abstract and the Decision Index and latency comparisons; readers interested in the training recipe will need the SFT and RLAF sections and the appendices.

Authors’ 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