Towards Foundation Decision Models: Drex
Large language models are usually asked to generate an answer token by token. That design is powerful, but it can be an awkward fit for production systems that need to choose among user-supplied options, return calibrated probabilities, and respond quickly. The Nace.AI Research Team’s paper, Towards Foundation Decision Models: Drex, presents Drex as an alternative: a family of models that scores the available options directly, without autoregressive decoding.
The problem with generating decisions
Many real applications do not need an essay. They need a classification, a yes-or-no judgment, a ranking, or a score over a scale. Prompting a generative model to produce such an answer introduces several practical problems. The output must be parsed, the model may produce invalid text, and generation adds latency and cost. Computing probabilities for arbitrary options can also require scoring each option separately, while constrained decoding can produce distributions influenced by tokenization, option length, or shared prefixes.
Drex treats the option set as part of the request. A request contains a state, such as a document, record, agent context, or optional image, plus one or more questions. Each question supplies its own options, and the model returns a probability distribution over exactly those options. The interface supports named choices, yes-or-no questions, and ordered score levels.
Two architectural strategies
Three models in the family—Drex DLM, Drex v1.5, and Drex v2.0—use pointer heads. The model encodes the state once, caches that representation, and processes each question as a separate branch. A query representation at the decision marker is compared with a representation for each option. One softmax then produces the distribution, with no generated answer sequence. Because the state is shared, these models are particularly suited to asking many questions about a long document.
Drex Max takes a different approach. It uses a dense 27-billion-parameter multimodal backbone, bidirectional attention in selected layers, and a readout head that maps the final hidden state to logits for up to 255 option codes. Max re-encodes the context for each question, so it gives up shared caching in exchange for a richer representation of the complete question and option list. The four models range from 8B parameters for DLM to 27B for Max; v2.0 has 35B total parameters with about 3B active.
The training recipe combines supervised fine-tuning with consistency losses, reinforcement learning from adjusted feedback (RLAF), and, for Max, image-encoder tuning. RLAF combines verified outcomes with a teacher or judge distribution while using proper scoring objectives intended to reward both accuracy and calibrated uncertainty. The authors emphasize that most comparisons are observational trade-offs rather than clean causal demonstrations, although they conducted a controlled ablation for Max.
What the evaluations show
The headline result is Drex Max’s performance on Decision Index 0.4.1. The authors report a public-suite score of 67.52, compared with 60.70 for Jev 1.14.0, and a full score of 66.60 that includes private tests. On the independently maintained Artificial Analysis Decision Index, Max scores 63.7; the paper describes it as near the top of the decision-model group, while noting that the small score differences do not justify a definitive ranking without confidence intervals.
On Decision Bench v4, Max reaches 94.0% accuracy when images are supplied through its image encoder. That is statistically indistinguishable from GPT-6 Sol and d3-turbo in the authors’ paired comparison, but 1.6 percentage points below Claude Opus 5.5. Max is much faster in the reported setup: its median end-to-end time is 0.66 seconds, versus 2.05 seconds for GPT-6 Sol and 4.33 seconds for Claude Opus 5.5.
Images are an important part of the result. On 60 image-only questions, Max achieves 98.3% accuracy with the actual images but 63.3% when given only text renderings. The authors interpret this as evidence that the vision stage materially improves image grounding, especially for extraction and document-oriented tasks. It is not evidence that the model matches frontier systems on all visual reasoning.
The models also provide useful confidence signals. Across about 230,000 decisions, deployed expected calibration error ranges from 1.37% for v2.0 to 4.54% for v1.5, with Max at 3.61%. A temperature fitted on held-out Decision Index data improves these figures, but calibration varies by domain and option count. Max is notably over-confident on questions with six to 20 options. Selective use helps: retaining the 50% of Max’s decisions with the highest confidence raises reported accuracy to 99.3% on that subset.
Order sensitivity and training effects
Reordering options changes Max’s top choice on 4.3% of the studied decisions under a random permutation. The rate rises to 13.3% for decisions with six or more options. Accuracy changes little, and the unstable cases are generally low-confidence cases. Consistency losses reduce Max’s measured flip rate from 6.3% to 4.2% in the reported ablation without materially changing benchmark accuracy. RLAF adds two points to the Decision Index score relative to supervised fine-tuning alone in that experiment.
Limitations and practical implications
Drex Max remains weaker than frontier chat models on difficult closed-book knowledge tests such as Humanity’s Last Exam and GPQA Diamond. The authors also cannot isolate whether Max’s gains come from its larger backbone, attention pattern, readout head, training data, or training procedure. The option-order study is relatively small, and benchmark contamination through teacher models cannot be ruled out completely. Long uncached contexts remain expensive, while Max’s cost grows with the number of questions because it cannot share the encoded state.
For practitioners, Drex is most compelling where the choices are known, structured, and numerous: routing, retrieval screening, tool selection, policy checks, document classification, and agent evaluation. Its direct probabilities make thresholding, abstention, auditing, and human review easier than parsing free-form model output. But the paper’s results support a cautious deployment pattern: tune thresholds on representative labeled traffic, monitor calibration by workload, log the full distribution and model version, and route low-confidence or consequential decisions to another system or a human. Drex is a strong argument for treating decision inference as its own model category—not as a smaller version of text generation.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.