Drex Max posts 67.52 on Decision Index and replies in 183 ms

Towards Foundation Decision Models: Drex

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

cs.CL

2026-11-05

Drex Max (27B) returns option probabilities with no autoregressive decode: 67.52 on Decision Index 0.4.1 vs 60.70 for Jev 1.14.0, in 183 ms.

What problem this solves

Classification, scoring, and routing do not need generated text. Long chain-of-thought still has to be parsed, latency tracks decoding, and a reworded question can change the answer. A classification head freezes its labels at training time. Scoring each option with its own forward pass, then renormalizing, multiplies cost by the number of options.

A decision model lets the caller supply the options. A request is a state, meaning a document, record, or agent context, images optional, plus one or more questions. One forward pass returns a distribution over those options. Nothing is decoded. The interface is Jev's System One, and the pointer models sit on the open Kev stack. Drex ships four weights: diffusion DLM at 8B, v1.5 at 9B, v2.0 at 35B total with 3B active, and Max at 27B, trained from Qwen3.8-27B.

Method

All three question types return a distribution. choice allows up to 255 named options. noul is yes/no and returns p(yes). score treats integer levels as options and also returns the expected level. Confidence follows System One: how far the top probability sits above a uniform guess. Two properties are architectural. Probability mass stays on the supplied options, and questions in one request do not affect each other. Order robustness, calibration, and vision are learned.

DLM, v1.5, and v2.0 use a pointer head. The state is encoded once and cached. Each question is a causal block that cannot see the others, so the distribution matches whether the question is asked alone, batched, or replayed from cache. The query is the hidden state at <decide>. The key is the hidden state at each option's closing </opt>. A scaled dot product and a softmax score the options. The head is about 2.1M parameters on v1.5 and 1.0M on v2.0. v2.0's backbone is Qwen3.6-35B-A3B. v1.5 and v2.0 mix in Gated DeltaNet, recurrent linear attention, and cannot take an arbitrary mask. DLM attends in both directions inside the state.

Max uses a readout head, following the public pplx-decider pattern. Of 64 layers, 48 still run left to right. Every fourth layer is full attention with the causal mask removed, so earlier tokens can see the options. Each question is its own sequence. Options carry single-token codes, and a bias-free linear layer reads logits from <ans>, up to 255 codes. Temperature 1.009 does not change the top option. Bidirectional attention makes the state depend on the question, so it cannot be cached across questions. Median latency for one request at launch was 183 ms for Max, 151 ms for v2.0, and 145 ms for v1.5. Many questions over a long document still favor the pointer models. On Max, binding a code to option text is learned. Permute the options and which logit belongs to which text changes.

Supervised training uses cross-entropy. Score questions add a squared Cramér distance, a penalty on the cumulative distribution. Two consistency terms penalize a distribution shift after the options are permuted, or after sentences or fields in the state are shuffled. Questions with three or more options are shuffled every time they are seen, and a NONE option is inserted at random. RLAF drops one-hot labels, which train over-confidence on uncertain items. Pointer models adjust a teacher distribution with a gold label, a program execution, or a game engine, then differentiate a log score, a spherical score, and the Cramér term. Max's reward mixes a hit on the verified answer with the judge's log probability, at ρ=0.2 and η=0.05, plus a KL pull back to the supervised checkpoint. The judge is the generative form of the same 27B at high reasoning effort, not a stronger model. The v1.5 text mix has 1,150,460 records. Decision Index and JevBench eval items are removed. Training splits of datasets used as public benchmarks stay in. Paraphrases, and benchmarks the teacher itself saw, are not filtered. How often the judge overrides the verified answer was not measured.

Results

Decision Index 0.4.1 reports chance-corrected skill, where 0 is uniform guessing and 100 is perfect. On the public board of 22 October 2026, Max scores 67.52, v2.0 61.25, v1.5 58.08, Jev 1.14.0 60.70, and GPT-6 Luna Decisions 47.78. Full scores that include private tests are 66.60 for Max (0.9 below its public score), 52.76 for v2.0 (down 8.5), and 45.78 for v1.5 (down 12.3). On new-domain private items Max scores 56.45 against 39.22 for v2.0. Max leads Jev's full score of 60.24 by 6.4 points. On about 230k choice questions Max is 1.86 points more accurate than v2.0. GSM8K moves from 66.8 to 83.2. POP909-CL, which depends on long structured options, falls from 79.5 to 55.6. On samples of about 100 requests each, HLE is 21% for Max, 39% for GPT-6 Sol, and 56% for Claude Opus 5.5. GPQA Diamond is 55%, 90%, and 89%.

Decision Bench v4 has 1,071 questions.

ModelAccuracyImage-onlyMedian time
Drex Max (image encoder)94.0%98.3%0.66 s
Drex Max (text rendering)92.1%63.3%0.66 s
d3-turbo93.7%98.3%0.72 s
GPT-6 Sol94.7%100%2.05 s
Claude Opus 5.595.6%100%4.33 s
Jev 1.1492.3%66.7%0.51 s
Drex v2.087.9%68.3%0.38 s

Max's 94.0% has a Wilson interval of [92.4, 95.3]. It is statistically tied with d3-turbo and GPT-6 Sol, and 1.6 points behind Claude. Bench ECE is 1.40%, against 1.18% for d3-turbo. A paired test versus Jev gives p=0.048, and the two models tie on text. The gap is the image questions. On the design category, accuracy rises from 26.7% with a text rendering to 96.7% with the image encoder.

A separate calibration slice of about 115k questions gives Max 87.9% accuracy, 91.2% mean confidence, and 3.61% ECE. Refitting temperature to 1.35 cuts ECE to 1.35%. v2.0's ECE on that slice is 1.37%, the lowest of the group. The most confident half of Max's decisions reaches 99.3%. Language is 56% of the items and has an ECE of 0.5%, which pulls the pooled figure down. On questions with 6 to 20 options, Max's ECE is 8.4% at 71.7% accuracy, while v2.0's is 2.0%.

On 1,591 decisions, a random permutation flips the top option for 4.3% of Max's answers and 4.3% of Jev's. v2.0 flips 5.9%. Binary items flip 1.0% of the time for Max. With six or more options every model flips 13% to 16%, and that bin has only 173 decisions. Abstaining when the top option is unstable across orders is no more accurate than abstaining on the same number of low-confidence items, 87.2% versus 87.1%. In the Max ablation, RLAF adds 2.0 Decision Index points, and the consistency losses cut the flip rate from 6.3% to 4.2%.

On the Artificial Analysis board, retrieved 29 October 2026 and not rerun by the authors, Max scores 63.7. The three leading decision models score 65.3, 65.0, and 64.7, with no confidence intervals. HLE multiple choice is 16.0 for Max and 70.2 for Claude. Chat models cost about 44 to 1,100 times more per request. On vision, InfographicVQA is 98.8% and multi-step MMMU-Pro is 53.6%. Against GPT-6 Luna Decisions, normalized accuracy on 11 vision benchmarks is 77.1 versus 63.9.

On decisions with a settled label, Max sits next to Claude. Knowledge items and multi-step visual reasoning stay well behind.

Why it matters

Routing, review, and picking from a candidate list skip generation. The caller passes the options in. Three options and 150 options share one head. The probabilities are already an audit record, and low confidence can go to a person. Max's most confident half reaches 99.3%. A typical request at launch returns in 183 ms, against about 2.0 s for GPT-6 Sol and 3.5 s for Claude Opus 5.5. Long documents with many questions fit v2.0 or v1.5. Images, or a higher ceiling, fit Max.

The lead over Jev holds on both the public index and the full score. The gap to Claude and Sol on Decision Bench is within 1.6 points. HLE and GPQA are a different tier. The pointer and the readout follow Kev and pplx-decider. What is new is RLAF, the consistency losses, and the vision stage. Licenses are not one license. Max, v2.0, and v1.5 use Nace.AI Open RAIL-M. DLM is CC BY-NC 4.0, so commercial use is excluded. Repository code is Apache-2.0.

Limitations

The shipped temperature, 1.009, was fit on held-out training records. Decision Index wants 1.35, and Max stays over-confident inside individual domains. One global temperature cannot fix errors that point in opposite directions. After a refit, GPT-6 Luna Decisions drops NLL from 0.760 to 0.679 while ECE rises from 4.87% to 6.57%. A narrow application that thresholds probability should refit temperature on labels from that application.

The judge can override a verified answer. With ρ at 0.2, a judge that places enough mass on a wrong option makes that wrong option optimal. That can sharpen the distribution, and it can copy the judge's bias. Neither rate was measured. Most training records are distilled from stronger teachers, and paraphrased benchmark items are not filtered. Outside one ablation, the comparisons are associations. How much of Max's gain comes from bidirectional attention, the readout, the data, or RLAF is unseparated. Tests were not corrected for multiple comparisons, so p=0.048 and p=0.046 are suggestive. Option-isolation masks were not tested on the DeltaNet backbones. On private new domains the pointer models drop much further than Max, and the cause is not separated. The same Max weights run with the original causal mask do not reproduce the trained scores.

Terms

Source

What people are saying

Related papers

All paper explainers