Towards Foundation Decision Models: Drex

Abstract: Autoregressive approaches with LLMs, 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 paper introduces a new model family, Drex, designed for dynamic decision-making by directly outputting calibrated probability distributions over user-specified options, enhancing choice accuracy and stability over conventional classification.
• Evaluated across a range of metrics and boosts benchmark performance, achieving a 67.52 public Decision Index score and 94.0% accuracy on Decision Bench with image inputs, maintaining real-world application.
• Drex v2.0 features question independence, high reliability, and low latency, while Drex scalable alternatives offer multimodal capabilities and improve decision performance across extensive validation frameworks.

Problem formulation and research objectives

“Towards Foundation Decision Models: Drex” (2611.07854) studies decision models as a distinct inference paradigm for applications in which the required output is a probability distribution over a user-specified set of options. Rather than generating an answer autoregressively and parsing text, Drex receives a state consisting of text, structured records, agent context, and, for Drex Max, images; it then evaluates one or more questions over that state. Each question may be a choice among arbitrary options, a binary yes/no judgment, or a score over an ordered integer scale.

This formulation addresses a structural limitation of conventional classification heads. Their label space is fixed during training, whereas decision-model options are supplied dynamically at inference time. Constrained decoding can guarantee that a generated answer belongs to the option set, but it still requires autoregressive computation and can produce probabilities distorted by tokenization, option length, shared prefixes, and per-token renormalization. Scoring each option independently avoids some of these problems but requires one forward pass per option. Drex instead trains a direct readout over the supplied options, returning probabilities without decoding.

The paper evaluates four properties required for production decision models:

• calibrated probabilities suitable for thresholding and selective prediction;
• robustness to arbitrary option ordering;
• independence between questions in the same request;
• low latency, particularly when many questions concern one long state.

Four research questions organize the study. The first concerns calibration across domains and option counts. The second measures sensitivity to option permutations. The third examines supervised fine-tuning (SFT), reinforcement learning from adjusted feedback (RLAF), consistency regularization, and vision tuning. The fourth compares cacheable pointer architectures with a larger, fully bidirectional multimodal readout architecture.

The Drex model family

The family contains three pointer models and one readout-head model:

Model	Backbone scale	Active parameters	Head	Modalities	Shared state
Drex DLM	8B	8B	256-dimensional pointer	Text	Yes
Drex v1.5	9B	9B	256-dimensional pointer	Text	Yes
Drex v2.0	35B MoE	3B	256-dimensional pointer	Text	Yes
Drex Max	27B dense	27B	255-way readout	Text and images	No

The pointer models serialize the state followed by independent question blocks. A block-causal or row-wise computation ensures that each question can attend to the state and its own instruction and options, but not to other questions. The state is therefore encoded once and cached. At a decision marker, the model constructs a query representation; at the end of each option, it constructs a key representation. A scaled dot product between the query and each option key produces logits, and a single softmax yields the distribution over exactly the options supplied by the caller.

This design provides two guarantees by construction. First, the output support is closed: probability mass exists only over the supplied options. Second, the answer to one question is independent of which other questions are included in the request, subject to floating-point effects. The architecture does not, however, guarantee option-order invariance, because the hidden representation of an option can depend on the preceding option sequence in hybrid backbones.

Drex Max adopts a different trade-off. It is initialized from a dense Qwen3.8-27B backbone with 64 layers, including Gated DeltaNet layers and full-attention layers. Causal masking is removed in the full-attention layers, allowing tokens near the beginning of the sequence to attend to later instructions and options. Each question is encoded as a separate sequence, so the state representation is question-dependent and cannot be shared across questions.

The answer is read from a final answer marker through a bias-free linear layer with 255 output logits. Options are prefixed with single-token codes such as A, B, and C; only the logits corresponding to the supplied options are normalized. This approach permits full bidirectional interaction among the state, instruction, and options, but it introduces a symbol-binding problem: the model must associate each option code with the text that follows it. Unlike the pointer architecture, option-order robustness is learned rather than enforced.

Training methodology

The text training mixture contains 1,150,460 verified decision records across eleven domains, including language understanding, tools and routing, knowledge, code and mathematics, retrieval, finance, policy, commerce, uncertainty judgments, media, and games. Labels are derived from gold answers, program execution, game engines, or other verification procedures. Teacher-generated distributions are used in addition to one-hot labels where available.

SFT combines cross-entropy with a squared Cramér distance for ordered score questions. Two consistency terms target behaviors that are not guaranteed by the architecture. A permutation-consistency term penalizes changes in the predicted distribution when options are shuffled, while a state-shuffle term penalizes changes when sentences or structured fields in the state are reordered. Training also introduces distractor or “none” options in some examples, reducing the tendency to force a choice when no candidate is appropriate.

RLAF replaces a purely categorical target with a distribution adjusted using both verified outcomes and teacher feedback. For pointer models, the adjusted distribution is optimized through proper scoring rules combining logarithmic, spherical, and ordered-outcome objectives. Drex Max uses a KL-regularized reward formulation: verified outcomes and a generative judge contribute to the reward, while a KL penalty constrains the model relative to its SFT checkpoint. The judge is the generative form of the same base model rather than an independently stronger system. The paper explicitly notes that a sufficiently confident but incorrect judge can override a verified answer under the selected reward coefficients, although the frequency of this event is not measured.

Drex Max adds a third vision-SFT stage that trains the image encoder and its projection into the language backbone. This stage is intended to preserve text performance while enabling decisions grounded in image tokens rather than captions or OCR-derived descriptions.

Accuracy and benchmark performance

Drex Max is strongest on the Decision Index 0.4.1 evaluation. On the public suite, it achieves a chance-corrected skill score of 67.52, compared with 61.25 for Drex v2.0, 58.08 for Drex v1.5, and 60.70 for Jev 1.14.0. Its full score, incorporating private same-skill and new-domain tests, is 66.60. The corresponding full scores for v2.0 and v1.5 are 52.76 and 45.78. The difference is especially pronounced on new-domain tests, where Max scores 56.45, compared with 39.22 for v2.0 and 30.82 for v1.5.

The public-domain gains are broad rather than concentrated in one benchmark. Max leads in Knowledge, Language, Retrieval, and Tools, and is only marginally behind Drex v2.0 in Arts. In the authors’ paired evaluation, Max scores higher than v2.0 on 28 of 37 benchmarks and is more accurate by 1.86 percentage points over approximately 230,000 shared choice decisions. Large improvements occur on GSM8K, BBH, ACOS, WinoGrande, MuSR, and the home-appliance simulator.

The gains are not universal. Max performs substantially worse than v2.0 on POP909-CL, falling from 79.5 to 55.6, and on Habermas Machine, falling from 62.4 to 50.5. Both tasks involve long structured sequences in the Arts domain. The paper hypothesizes that the readout design’s requirement to bind option codes to long option descriptions may be disadvantageous, but the architecture and training changes are confounded, so this explanation is not established causally.

On Decision Bench v4, Max with image inputs reaches 94.0% accuracy over 1,071 questions. This is statistically tied with d3-turbo and GPT-6 Sol, and is 1.6 points below Claude Opus 5.5. It exceeds v2.0 and v1.5 by 6.2 and 7.2 points, respectively. On text-only questions, Max reaches 93.4%. Its strongest multimodal result is on image-only questions: image input raises accuracy from 63.3% with text renderings to 98.3% with the image encoder.

Evaluation	Drex Max	Relevant comparison
Decision Index public skill	67.52	Jev 1.14: 60.70
Decision Index full skill	66.60	Jev 1.14: 60.24
Decision Bench v4	94.0%	Claude Opus 5.5: 95.6%
Image-only Decision Bench	98.3%	Text rendering: 63.3%
JevBench-231	91.8%	d3-turbo: 90.9%; GPT-6 Luna: 91.3%
DecideBench item accuracy	98.5%	Jev: 98.0%
DecideBench pair accuracy	97.0%	Jev: 96.0%

On JevBench-231, Max answers 212 of 231 items correctly, or 91.8%. Its advantage over v2.0 is concentrated in the hard tier, where accuracy rises from 73.9% to 82.9%. On DecideBench, which uses minimal pairs designed to test whether a model reacts to a small change in the state, Max attains 97.0% pair accuracy. It also obtains the lowest reported ECE on that benchmark, although the estimate is based on only 400 questions.

The results are weaker on closed-book knowledge than on real-data decision tasks. On a fixed 100-question sample, Max scores 21% on Humanity’s Last Exam and 55% on GPQA Diamond, compared with 39% and 90% for GPT-6 Sol and 56% and 89% for Claude Opus 5.5. The paper therefore does not support the claim that a discriminative decision head closes the capability gap on knowledge-intensive reasoning. A single-pass decision model remains disadvantaged relative to systems that can perform extended chain-of-thought reasoning or tool use.

Calibration and selective prediction

Calibration is evaluated on approximately 230,000 gold-labeled choice decisions per model from Decision Index 0.4.1. Drex v2.0 is the best calibrated at its deployed temperature, with an ECE of 1.37% and an accuracy of 86.0%. Drex Max is the most accurate, at 87.9%, but is over-confident: its mean confidence is 91.2% and its ECE is 3.61%. Drex v1.5 has an ECE of 4.54%.

A temperature fitted on a held-out subset of Decision Index improves ECE for the Drex models to 0.79% for v2.0, 1.17% for v1.5, and 1.35% for Max. For Max, the fitted temperature is 1.35, substantially higher than its deployed value of 1.009. This discrepancy indicates a distribution shift between the training-derived calibration set and Decision Index. It also demonstrates that a global deployed temperature is not a reliable substitute for calibration monitoring on the target workload.

Calibration varies sharply by domain. Max has ECE values of 7.2% to 9.2% in Knowledge, Retrieval, Tools, and Arts, despite an aggregate ECE of 3.61%; Language contributes 56% of the decisions and is nearly calibrated. Option count is similarly consequential: Max has an ECE of 8.4% on questions with 6–20 options, where accuracy is 71.7% and mean confidence is 80.1%. Consequently, applications that use confidence thresholds should tune them on labeled data from the specific domain and option-count distribution.

The confidence estimates nevertheless support effective selective prediction. Retaining the 80% of Max’s decisions with the highest confidence raises accuracy from 87.9% to 96.1%; retaining the top 50% raises it to 99.3%. This result directly supports routing low-confidence cases to human review or a more capable system, provided that the deployment distribution resembles the calibration data or is monitored continuously.

Figure 1: Calibration curves, selective accuracy, and domain-specific ECE show strong aggregate calibration but substantial over-confidence for Drex Max in several domains and option-count regimes.

Option-order robustness

The paper evaluates 1,591 decisions under original, reversed, and random option orderings. Under a random permutation, the top option changes for 4.3% of Max’s decisions, 5.9% of v2.0’s, 5.3% of v1.5’s, and 4.3% of Jev 1.14’s. Reversal produces somewhat higher flip rates, between 5.7% and 8.4%. Max is therefore no less stable than the comparison model despite using a position-indexed readout head.

The effect on accuracy is small: Max changes from 83.6% under the original ordering to 83.4% under random permutation. The instability is concentrated in difficult decisions. For Max, the flip rate is 1.0% for binary questions, 7.2% for questions with three to five options, and 13.3% for questions with at least six options. These are usually low-confidence decisions where several options have similar probability.

Order instability is diagnostically useful but not more useful than confidence alone. Max’s accuracy is 87.2% on decisions stable across all three orderings and only 37.6% on unstable decisions. However, abstaining on unstable decisions performs nearly identically to abstaining on an equal number of the lowest-confidence decisions: 87.2% versus 87.1% accuracy at the same coverage. Averaging predictions over the three orderings improves Max’s log loss by 0.014 nats but does not materially improve accuracy, and it increases inference cost.

Figure 2: Option permutation primarily changes low-confidence multi-option decisions; the resulting accuracy differences remain below 1.2 percentage points.

The controlled Max ablation confirms the role of consistency regularization. Removing the permutation and state-shuffle terms increases the random-permutation flip rate from 4.2% to 6.3%, and the flip rate for questions with at least six options from 13.0% to 17.8%. Decision Index accuracy remains essentially unchanged. Thus, the consistency losses improve behavioral invariance rather than core task accuracy.

Training-stage analysis

The checkpoint analysis of Drex v2.0 shows that most of the improvement occurs during SFT. The Decision Index 0.2.1 skill score rises from 58.27 at an early SFT checkpoint to 60.02 near the end of SFT, then largely plateaus. The specialization stage adds approximately 0.6 points at its best checkpoint, primarily through a 1.8-point Retrieval gain, and raises JevBench accuracy from 84.0% to a peak of 87.0%. RLAF adds only 0.2 points to the Decision Index score in this sequential training trajectory, with small gains in Tools and Arts.

The Max ablation provides stronger evidence for RLAF because its variants share the same backbone, data, and schedule. Relative to SFT alone, RLAF increases the Decision Index score by 2.0 points, raises accuracy by 1.5 percentage points, and reduces ECE by 0.9 percentage points. This is the clearest controlled training result in the paper.

The vision stage has no measurable cost on text inputs: text Decision Index score, accuracy, and option flip rate remain nearly unchanged. Its multimodal effect is substantial, however, improving image-only Decision Bench accuracy from 63.3% to 98.3%. The separation between text preservation and image grounding is one of the strongest empirical results supporting the three-stage training design.

These results should not be interpreted as isolating the independent causal contribution of every component. Max differs from the pointer models simultaneously in backbone size, dense versus sparse computation, attention masking, readout mechanism, training schedule, and modality. The observed Max–v2.0 gap therefore establishes a system-level performance difference, not the effect of any single architectural decision.

Inference cost and deployment

Drex’s operational advantage comes from replacing autoregressive decoding with a forward pass and a direct softmax over options. Pointer models encode the state once and reuse it across questions. A subsequent question about a cached 114k–117k-token state takes 21.6 ms, faster than processing a new uncached 2k-token state, which takes 35.6 ms. Shared prefill yields approximately 6–8x speedups at 8k tokens and 11–12x at 32k tokens with sixteen questions.

At the reported deployment configuration, median single-request latency is 145 ms for v1.5, 151 ms for v2.0, and 183 ms for Max. The corresponding commercial frontier baselines require approximately 2.0 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5. Reported prices per 1,000 decisions are $0.04 for v1.5,$0.06 for v2.0, $0.06 for Max,$3.33 for GPT-6 Sol, and $10.17 for Claude Opus 5.5.

The latency claims require qualification. They include network-dependent measurements in some comparisons, and long uncached contexts remain expensive. For v1.5, a 131,072-token state takes approximately 2.7 seconds, exceeding the stated two-second target. Max also loses the pointer models’ principal advantage for multi-question workloads because it must re-encode the state for every question. Low-precision kernels and batch-shape-dependent tiling can produce small probability changes, so the paper recommends operational margins around threshold decisions.

Limitations and open questions

The principal limitation is architectural confounding. Max’s gains cannot be attributed separately to its 27B dense backbone, non-causal full-attention layers, readout head, multimodal training, or modified RLAF procedure. The paper identifies a direct experiment that remains outstanding: train pointer and readout variants on the same Qwen3.8-27B backbone and data, with and without non-causal attention, and compare them under matched optimization.

The calibration analysis also leaves deployment-specific uncertainty. Aggregate ECE conceals large domain-level errors, and the temperature fitted on Decision Index does not necessarily transfer to another application. The option-order study is based on a relatively small, filtered sample and excludes structured options rejected by public APIs. Its conclusions should therefore be treated as estimates for the tested workload rather than general invariance guarantees.

The RLAF judge introduces an additional unresolved issue. Because the judge can override a verified answer under the stated reward formulation, the training procedure may propagate judge biases or induce over-confidence. The paper does not measure the frequency of verified-answer overrides, nor does it quantify how much of Max’s calibration error is attributable to judge-derived feedback.

Finally, the benchmark results do not establish broad superiority over frontier generative systems. Max is competitive on real-data decision benchmarks, multimodal extraction, routing, and low-latency classification, but it remains substantially weaker on difficult closed-book knowledge and extended reasoning. The specific question left open is whether allowing a decision model to use retrieval, tools, or an intermediate reasoning process before applying its non-autoregressive readout can improve those tasks without eliminating its latency and calibration advantages.

Conclusion

Drex presents a coherent family of open decision models that directly produce calibrated distributions over dynamically supplied options. Its pointer models provide question independence, shared-state caching, and low-latency inference; Drex Max sacrifices state reuse for a larger dense multimodal backbone and bidirectional attention. The strongest system-level result is Max’s 67.52 public Decision Index score and 94.0% Decision Bench accuracy with image inputs, combined with substantially lower latency than frontier chat models.

The evidence also defines the boundaries of the approach. RLAF improves accuracy and calibration, consistency losses improve option-order stability, and image tuning produces a large multimodal gain without degrading text performance. However, Max remains over-confident in several domains, is less reliable on many-option decisions, exhibits residual order sensitivity, and trails frontier models on difficult knowledge-intensive tasks. Drex therefore establishes a technically credible design point for high-throughput probabilistic decision inference, while leaving the causal contribution of its architectural components and its performance under extended reasoning unresolved.