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
Figures & tables
Figure 1: Drex pointer architecture (DLM, v1.5, and v2.0, not including Drex Max, which is shown in Figure 2 ). The state and the questions (left) are tokenized into one sequence that consists of a state segment S followed by one question block bk for each question. Each question block is a contiguous block of text that contains the instruction, the options, and a final <decide> token, in this order; each option is enclosed between an opening tag <opt> and a closing tag </opt> . The backbone fθ encodes the sequence under the block-causal pattern shown inside it ( Eq. 17 ): every token of a question block can attend to all tokens of the state and to the preceding tokens of the same question block, but not to any token of another question block. The pointer head turns the hidden state at <decide> into a query q and each </opt> state into a key kj , and returns a softmax over exactly the supplied options (right). The example outputs are consistent with Eq. 1 : (0.62−\nicefrac13)/(1−\nicefrac13)=0.43 and s^=0.83+2×0.12=1.07 .
Figure 2: Two ways to score the options of a question. Each small square is one token, colored by its type as in Figure 1 ; the legend at the bottom lists the token types. (a) The pointer models (DLM, v1.5, and v2.0) encode one sequence per request: the state S , followed by one question block per question. Each question block attends to the state but not to other question blocks, so the state is encoded once and shared by all questions. The pointer head scores each option as the scaled dot product of the hidden state at the option’s </opt> token (the key) and the hidden state at the <decide> token of its question block (the query). (b) Drex Max encodes each question as its own sequence: the image tokens that an image encoder (ViT) produces from the input images, the state S , the instruction, the options, and a final answer marker, the single token <ans> . Each option is written after its option code ( A , B , …), which is a single token, and a colon. The 48 Gated DeltaNet layers of the backbone process the sequence from left to right, and its 16 full-attention layers (every fourth layer) attend in both directions. A readout head, i.e., a bias-free linear layer, maps the final hidden state at <ans> to one logit per option code (255 in total), and a softmax over the logits of the K supplied options gives their probabilities. The pointer models trade accuracy for a shared, cacheable state; Drex Max trades the cache for full bidirectional context in every question.
Figure 3: Post-training pipeline of the pointer models (shown for Drex v1.5). Stage 1 fine-tunes every weight in the backbone and a randomly initialized pointer head by optimizing the supervised objective of Eq. 9 . Stage 2 (RLAF) replaces one-hot labels with an adjusted target q~ , which is constructed from a teacher distribution and the verified outcome. RLAF then maximizes the expected reward R of Eq. 10 , which is largest when the predicted distribution equals q~ . Drex Max is trained with full-parameter SFT (with the same consistency losses), then the low-rank variant of RLAF as described in Section 6.5 , then SFT of the image encoder ( Appendix D ).
Figure 4: Training data of Drex v1.5: 1,150,460 verified text decision records over eleven domains. Each bar shows the number of records in the respective domain.
Private
Public
Public per-domain skill
Model
Active / total
Full
Same
New
Index
Raw
Know.
Lang.
Retr.
Tools
Arts
Drex Max
27B / 27B
66.60
63.08
56.45
67.52
75.40
57.1
72.2
68.8
81.9
53.4
Drex v2.0
3B / 35B
52.76
54.30
39.22
61.25
70.55
47.9
65.3
63.0
76.5
53.9
Drex v1.5
9B / 9B
45.78
49.65
30.82
58.08
68.08
44.6
60.4
61.7
75.0
48.8
Jev 1.14.0
N/A
60.24
59.28
48.26
60.70
69.98
57.1
61.2
63.5
73.8
39.2
Jev 1.13.0
N/A
59.22
58.04
47.72
57.96
68.55
53.9
59.2
55.4
75.1
39.1
Table 1: Decision Index 0.4.1 (leaderboard of 22 October 2026). Same : subscore computed on the same-skill portion of the private subset; New : subscore computed on the new-domain portion of the private subset; Raw : the unweighted average accuracy on the public subset of 37 benchmarks, which is shared with version 0.3; Index : the average accuracy on the public subset weighted as per weights provided by Decision Index; the last five columns give the chance-corrected skill on the public subset in each of the five domains of benchmarks (Knowledge, Language, Retrieval, Tools, Arts). Best per column in bold.
Figure 5: Per-domain skill on Decision Index 0.4.1 (public subset, scores from the leaderboard, except for GPT-6 Luna Decisions, which we evaluated locally). Drex Max leads in every domain except Arts, where Drex v2.0 is 0.5 points higher, and Knowledge, where it ties with Jev 1.14.0 at 57.1.
Accuracy by modality (%)
Model
Accuracy (%)
Text
Text+Img
Image
ECE (%)
Log loss
Median time per request (s)
Drex Max (image encoder)
94.0 [92.4, 95.3]
93.4
100.0
98.3
1.40
0.159
0.66
Drex Max (text rendering)
92.1 [90.3, 93.5]
93.4
100.0
63.3
1.51
0.194
0.66
Drex v2.0 †
87.9 [85.8, 89.7]
88.3
100.0
68.3
1.73
0.287
0.38
Drex v1.5 †
86.8 [84.7, 88.7]
86.9
98.4
73.3
4.77
0.339
0.26
Jev 1.14 †
92.3 [90.5, 93.7]
93.4
100.0
66.7
1.60
0.212
0.51
Table 2: Decision Bench v4 (1,071 decisions; consisting of 949 text, 62 text+image, and 60 image examples). Accuracy with 95% Wilson confidence intervals; expected calibration error (ECE) and log loss from the model’s probabilities; median end-to-end wall time per request from our client. Drex Max (text rendering) receives only the text rendering of each image.
Figure 6: Other decision benchmarks. One dot per model; Drex models are colored purple , and axes do not start at zero. The dashed line indicates the Jev 1.14 reference win rate of 50% in the game arena, and it indicates the “shorter transcript ⇒ success” heuristic (AUROC 0.709) on OSWorld-Verified. The game-arena win rates (bottom-left panel) come from a separate evaluation that was run when Drex Max was launched ( Table 17 in the Appendix) and have no confidence intervals. The results in the other three panels come from our own evaluations.
AA-Omniscience
Model
Index
Accuracy
Non-hallucination rate
HLE-MC
When2Call
WildGuard
Cost per 1M requests
Time per request (ms)
Claude Opus 5.5 (Max)
79.0
93.4
30.3
70.2
90.5
93.5
$51,295
3571
GPT-6.1 Sol (Low)
71.5
93.0
22.9
55.4
80.6
92.0
$2,048
2888
d1-ultra
65.3
31.6
84.5
15.0
92.6
95.4
$32
93
SubQ Snap 1.1
65.0
20.2
85.7
17.5
94.6
94.9
$30
190
Elephant 9B
64.7
11.6
93.2
18.3
94.8
93.3
$8
165
Table 3: Artificial Analysis Decision Index (retrieved 29 October 2026). The table lists two language models at their best setting, the three leading decision models, the Drex models, and the Jev, Decider and GPT-6 Luna Decisions models, out of a total of 35 entries on the leaderboard (an entry is one model with one setting, so some language models appear more than once). The columns under AA-Omniscience list the accuracy and non-hallucination rate (share of wrong answers that were abstentions) of each model for AA-Omniscience-Decision; the numbers in the HLE-MC and When2Call columns are accuracies, and those under WildGuard are average F1 scores on WildGuardTest. Cost is in US dollars per million requests and time is the average per request, both as reported by Artificial Analysis.
Deployed temperature
Decision Index-tuned temperature
Accuracy on top- k % (%)
Model
Accuracy (%)
ECE (%) [95% CI]
NLL
Brier
T⋆
ECE (%)
NLL
k=80
k=50
Drex Max
87.9
3.61 [3.42, 3.83]
0.360
0.172
1.35
1.35
0.345
96.1
99.3
Drex v2.0
86.0
1.37 [1.20, 1.56]
0.382
0.194
0.91
0.79
0.381
94.4
98.9
Drex v1.5
85.3
4.54 [4.27, 4.77]
0.429
0.207
1.48
1.17
0.398
94.0
99.3
GPT-6 Luna Decisions
76.4
4.87 [4.65, 5.15]
0.760
0.326
1.91
6.57
0.679
85.1
91.8
Table 4: Calibration on Decision Index 0.4.1 decisions (held-out half, ≈ 115k decisions per model). Accuracy does not depend on the temperature. Deployed temperature : the model as released ( Tdep , Section 6.6 ); Decision Index-tuned temperature : temperature T⋆ fitted on the other half of Decision Index. Mean confidence is the average probability of the top option. ECE is expected calibration error, and NLL is negative log loss. Lower ECE, NLL, and Brier score are better. The last two columns give the accuracy on the k% of decisions on which the model is most confident.
Figure 7: Calibration of the deployed models on the evaluation subset of Decision Index 0.4.1. (a) Calibration curves: the accuracy of the decisions in each bin (of predicted probabilities) versus the average predicted probability of the decisions that bin, showing only bins with at least 30 decisions; a perfectly calibrated model lies on the diagonal. (b) Precision@top- k %: the accuracy on the k % of decisions on which the model is most confident, for k=25,50,75,100 . (c) Expected calibration error (ECE) in each Decision Index domain for the three Drex models, at the deployed temperature (bars) and at the temperature T⋆ fitted on a subset of Decision Index (circles). GPT-6 Luna Decisions is not shown in (c) because we do not have its per-domain breakdown.
Accuracy (%)
Top-1 flip rate (%)
Flip under permutation by K (%)
Model
Original ordering
Reversed ordering
Permuted ordering
Ensemble of 3 orderings
Reversed ordering
Permuted ordering [95% CI]
Unstable
Mean total variation distance
K=2
3≤K≤5
K≥6
Drex Max
83.6
83.3
83.4
83.5
6.2
4.3 [3.3, 5.3]
7.4
0.037
1.0
7.2
13.3
Drex v2.0
81.0
79.8
80.6
81.1
8.0
5.9 [4.8, 7.1]
9.8
0.050
2.0
9.5
16.2
Drex v1.5
78.8
78.3
79.6
79.6
8.4
5.3 [4.2, 6.5]
9.2
0.047
2.0
7.8
15.6
Jev 1.14
79.3
78.9
79.0
79.5
5.7
4.3 [3.3, 5.4]
7.0
0.037
2.0
5.2
14.5
Table 5: Option-order robustness on 1,591 decisions. The table gives the accuracy (%) of each model under the original ordering of the options, under the reversed ordering, and under a random permutation. It also gives the accuracy of an ensemble over the three orderings: for each decision, we average the three probability distributions that the model returns under the three orderings, option by option, and take the option with the highest average probability. This accuracy is therefore not the mean of the three other accuracies. The flip rate (%) is the fraction of decisions whose top option differs from the top option under the original ordering, with a 95% bootstrap confidence interval for the random permutation. “Unstable” is the fraction of decisions whose top option is not the same under all three orderings. The mean total variation distance is computed between the distributions under the original ordering and under the random permutation. K is the number of options of a decision.
Figure 8: Option order. (a) The flip rate under a random permutation of the options, that is, the fraction of decisions whose top option changes, for decisions with K=2 , 3≤K≤5 , and K≥6 options. (b) The accuracy of each model under the original, reversed, and permuted orderings of the options, and the accuracy of the ensemble of the three orderings, which averages the three probability distributions that the model returns for a decision, option by option, and takes the option with the highest average probability ( Table 5 ). Under the reversed and permuted orderings and the ensemble, the accuracy of every model differs from its accuracy under the original ordering by at most 1.2 percentage points.
Domain (skill)
Checkpoint
Decision Index 0.2.1 (skill)
Knowledge
Language
Retrieval
Tools
Arts
JevBench accuracy (%)
Game arena score vs. Jev 1.13.0 (%)
Stage 1 SFT, step 3.6k
58.27
44.9
64.6
61.1
73.0
43.9
82.3
57.0 [51.6, 62.3] (256)
Stage 1 SFT, step 6.8k
60.02
47.5
65.3
61.2
75.5
48.0
84.4
60.9 [55.5, 66.2] (256)
Stage 1 SFT, step 9.0k (end)
60.01
47.4
65.6
61.1
75.5
47.9
84.0
57.6 [50.3, 64.6] (144)
Stage 2 specialization, step 175
60.38
47.2
65.8
62.7
75.2
48.4
87.0
60.5 [55.1, 65.6] (256)
Stage 2 specialization, step 250
60.65
47.7
65.9
62.9
75.5
48.9
85.7
60.2 [54.9, 65.2] (256)
Table 6: Evaluation of Drex v2.0 training checkpoints. Each row gives the results of one saved checkpoint of Drex v2.0, evaluated on all benchmarks of Decision Index 0.2.1, on JevBench-231, and in the game arena. For the Decision Index, we use its official evaluation code and report the chance-corrected skill, both overall and in each of its five domains. For JevBench-231, we report accuracy. The game arena consists of eight games from OpenSpiel in which the model plays against Jev 1.13.0, choosing each move from the set of legal moves ( Appendix G ). Its score is (wins+\nicefrac12draws)/games , which we report with a 95% bootstrap confidence interval, with the number of games shown in parentheses.
Figure 9: Drex v2.0 across training stages. Stage 1 (S1) is supervised fine-tuning, Stage 2 (S2) is the specialization stage ( Section 6.4 ), and Stage 3 (S3) is reinforcement learning from adjusted feedback (RLAF). Most of the improvement in the Decision Index 0.2.1 skill score occurs in Stage 1. Stage 2 mainly improves the skill in the Retrieval domain and the accuracy on JevBench-231, and Stage 3 gives small further gains in the skill in the Tools and Arts domains. In the game arena against Jev 1.13.0 ( Table 6 ), the 95% confidence intervals of the score overlap at every checkpoint.
Variant
Decision Index
Accuracy (%)
ECE (%)
T⋆
Flip (%)
Flip, K≥6 (%)
SFT only
64.9
86.5
4.4
1.45
5.3
14.1
SFT → RLAF, without ℓπ , ℓs
66.4
87.7
3.9
1.38
6.3
17.8
SFT → RLAF (before vision SFT)
66.9
88.0
3.5
1.34
4.2
13.0
Released: SFT → RLAF → vision SFT
66.8
87.9
3.6
1.35
4.3
13.3
Table 7: Controlled comparisons on Drex Max. Each row corresponds to one variant of Drex Max. For each variant, the table reports four evaluation results and one fitted parameter. The first result is the Decision Index 0.4.1 score on the public benchmarks. For the released model, this score differs from the score of 67.52 on the leaderboard ( Table 1 ) because of the variation between repeated evaluations of the same model. The second and third results are the accuracy and the ECE on the held-out half of the choice decisions of the Decision Index ( Section 8.4 ). The fourth result is the flip rate under a random permutation of the options, on all 1,591 decisions of our study of option-order robustness, and on the 173 of those decisions with K≥6 options ( Section 8.5 ). The fitted parameter is the temperature T⋆ that is fitted on the other half of the Decision Index decisions. All variants are evaluated on the same set of decisions.
Drex Max
Drex v2.0
Drex v1.5
Jev 1.14
Luna Dec.
GPT-6 Sol
Opus 5.5
Median latency
183 ms
151 ms
145 ms
196 ms
326 ms
∼ 2.0 s
∼ 3.5 s
Cost / 1k decisions
$0.06
$0.06
$0.04
$0.042
$0.10
$3.33
$10.17
Table 8: Latency and cost. The table shows the median latency of a single request and the published price per 1,000 decisions of each model in October 2026, when Drex Max was released. We measured these latencies from a test machine in Oregon, which includes 70–215 ms of network time between the test machine and the servers.
Change
Answers changed
Accuracy difference
Questions, GPU
Short-context build
79 / 5,966
+0.06 [ −0.05 , +0.18 ]
28,789, B200
Long-context build
2 / 127
+0.00 [ −0.41 , +0.43 ]
2,737 (204 docs), B200
FP8 vs. bf16
—
+0.00
1,934, B200
Graph padding + small-batch FP8
—
−0.34 [ −0.55 , −0.13 ]
11,932, H100
Kernel fusion (short contexts)
1 / 365
—
365, B200
H100 graph policy (self-hosting)
11 / 365
—
365, H100
Table 9: Accuracy cost of answer-changing optimizations (Drex v1.5). Each row compares one change with the path it replaced, on the same questions; rows use different baselines and do not add up. Accuracy differences are in percentage points with 95% paired bootstrap intervals, resampled by document. “Contract” questions are an internal labeled set that requires reasoning over long contracts. The last two rows are measured on a 365-question panel, where one decision is worth 0.27 points.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Type
User supplies
Model returns
choice
named options, K≤255 (key order = option order)
argmax key, pj for every key, cchoice
noul
a yes/no question; optional definitions of yes/no
p(yes)∈[0,1]
score
an integer scale, lowest level first
s^ , per-level pi , cscore
Appendix
Table 10: The three question types , with what the user supplies and what the model returns. Every type returns a probability distribution over options supplied by the user.
Figure 10: Attention masks for a request with a four-token state, a three-way choice question ( b1 ), and a noul question ( b2 ). Filled cells mark the query–key pairs that are allowed to attend, and lighter cells mark tokens in the question block that attend to the shared state. (a) Block-causal mask of Eq. 17 . (b) Drex DLM adds bidirectional attention within the state. (c) Optional option isolation, in which <decide> still attends to every option. Hybrid backbones implement (a) by processing each question as a separate causal sequence [S;bk] that starts from the cached state. The masks are generated by the code that implements them in the released models.
SFT (full parameters)
RLAF (low-rank)
SFT of the image encoder
Trainable parameters
all backbone weights and the readout head
LoRA adapters (0.4B) and the readout head
image encoder and projector (0.6B)
Records / questions
1.15M / 2.9M
180k / 420k
95k / 140k
Steps / batch (records)
9,000 / 256
1,200 / 128
1,500 / 64
Peak learning rate
1×10−5 (backbone), 1×10−3 (readout)
5×10−5 (adapters), 1×10−4 (readout)
2×10−5
Schedule
linear warm-up (3%), cosine decay
constant after warm-up
linear warm-up (3%), cosine decay
λπ / λs (consistency)
0.05 / 1.0
0.05 / 1.0
—
Appendix
Table 11: Drex Max training configuration of the three training stages ( Appendix D ). λπ and λs weight the consistency terms, and β , η , and ρ are the coefficients of Eqs. 12 and 13 .
Drex DLM
Drex v1.5
Drex v2.0
Drex Max
Backbone
Efficient-DLM-8B
MiMo-V2.6-Distill-Qwen-9B
Qwen3.6-35B-A3B
Qwen3.8-27B
Layer types
block-bidir. diffusion
3:1 DeltaNet : full attn.
3:1 hybrid, MoE MLP
3:1 DeltaNet : full attn.
Layers / hidden size d
—
32 / 4096
40 / 2048
64 / 5120
Experts (routed / total)
—
—
8 / 256
— (dense)
Parameters (active / total)
8B / 8B
9B / 9B
3B / 35B
27B / 27B
Decision head
256-d pointer
256-d pointer
256-d pointer
255-way readout head
Appendix
Table 12: The Drex family. All four models accept the same request format and return the same output format. DLM, v1.5, and v2.0 each have their own pointer head with 256-dimensional query and key projections ( Section 4 ), and Drex Max uses the readout head of Section 5 . Index values are chance-corrected Decision Index skill ( Section 7 ): the public index and the full score on the 0.4.1 leaderboard and, for Drex DLM, the index of version 0.2, which is not comparable with 0.4.1.
Domain
Records
Share
Example tasks
Language & reading
210,124
18.3%
NLI, paraphrase, sarcasm, contracts
Tools, routing & home
181,004
15.7%
when to call, tool + arguments, side-effect checks
Table 13: Domains of the Drex v1.5 training mix , with the number of records in each domain and examples of the decision tasks that it contains.
Figure 11: Per-benchmark change from Drex v2.0 to Drex Max on Decision Index 0.4.1 (chance-corrected skill, our evaluations), grouped by domain and sorted within each domain. Max scores higher on 28 of the 37 benchmarks, and its only two large drops are on POP909-CL and Habermas Machine.
All 1,071 rows
949 text-only rows
vs.
Δ [95% CI]
Discordant
p
Δ [95% CI]
p
Drex v2.0 †
+6.2 [+4.4, +7.9]
81 / 15
< 0.001
+5.1 [+3.3, +6.8]
< 0.001
Drex v1.5 †
+7.2 [+5.2, +9.2]
99 / 22
< 0.001
+6.4 [+4.3, +8.5]
< 0.001
Jev 1.14 †
+1.8 [+0.2, +3.5]
51 / 32
0.048
+0.0 [-1.6, +1.7]
1.000
d3-turbo
+0.3 [-0.7, +1.3]
16 / 13
0.711
+0.3 [-0.7, +1.5]
0.711
GPT-6 Sol
-0.7 [-2.2, +0.9]
36 / 43
0.500
-0.6 [-2.4, +1.2]
0.572
Appendix
Table 14: Paired comparisons on Decision Bench v4 : accuracy of Drex Max (images in image encoder) minus that of each other model, in points. The discordant counts are the number of rows that Max answers correctly and the other model incorrectly, followed by the number for the reverse case. The 95% intervals are bootstrap intervals over rows, and p is from an exact McNemar test.
Figure 12: Decision Bench v4 by category , sorted by Drex Max accuracy (images in image encoder); category sizes in parentheses. The frontier models’ advantage is concentrated in Product management and AI agents & evals, and Drex Max’s largest deficit is Trust & safety.
Model
Accuracy
95% CI
Easy (48)
Standard (72)
Hard (111)
p vs. Max
Drex Max
91.8 (212/231)
[87.5, 94.7]
100.0
100.0
82.9
—
Drex v2.0
86.1 (199/231)
[81.1, 90.0]
100.0
95.8
73.9
0.002
Drex v1.5
84.8 (196/231)
[79.7, 88.9]
100.0
94.4
72.1
< 0.001
d3-turbo
90.9 (210/231)
[86.5, 94.0]
100.0
100.0
81.1
0.791
GPT-6 Luna Decisions
91.3 (211/231)
[87.0, 94.3]
100.0
98.6
82.9
1.000
GPT-6 Sol
99.1 (229/231)
[96.9, 99.8]
100.0
100.0
98.2
< 0.001
Appendix
Table 15: JevBench-231 (our evaluations). Overall accuracy (%) with its 95% Wilson interval, accuracy on each difficulty tier (number of items in parentheses), and the p -value of an exact McNemar test vs. Drex Max.
Table 28
Drex Max
Drex v2.0
Drex v1.5
Domain
N
Acc.
Conf.
ECE
ECE T⋆
Acc.
Conf.
ECE
ECE T⋆
Acc.
Conf.
ECE
ECE T⋆
Knowledge
24,266
75.7
82.0
7.3
4.4
70.8
72.1
2.8
3.7
67.4
77.7
10.5
4.7
Language
64,364
95.9
96.4
0.5
1.3
95.4
92.2
3.2
2.4
95.8
97.0
1.2
0.9
Retrieval
12,087
81.6
89.8
8.2
4.6
76.8
79.0
3.3
4.2
77.7
88.1
10.6
5.0
Tools
5,816
85.9
95.0
9.2
7.0
85.7
90.7
5.0
6.0
85.0
95.1
10.1
6.8
Arts
8,453
71.8
77.2
7.2
6.3
71.5
66.6
5.0
3.3
68.4
69.2
2.9
6.4
Appendix
Table 18: Calibration by Decision Index domain , on the held-out half of the decisions ( Section 8.4 ). For each model, the table gives the accuracy, the mean confidence (the probability of the top option), the ECE of the released model, and the ECE after rescaling with the single temperature T⋆ of Eq. 14 , fitted on held-out data (all in %).
Drex Max
Drex v2.0
Drex v1.5
Options
N
Acc.
Conf.
ECE
Acc.
Conf.
ECE
Acc.
Conf.
ECE
K=2
163,148
92.0
95.4
3.4
90.4
89.9
0.6
90.7
94.6
3.9
3≤K≤5
36,953
84.0
87.1
3.2
81.5
78.2
3.4
80.3
84.2
3.8
6≤K≤20
15,155
71.7
80.1
8.4
67.8
66.0
2.0
56.9
70.5
13.6
K>20
15,061
68.6
66.3
4.0
67.4
61.6
5.8
68.2
71.1
3.5
Appendix
Table 19: Calibration by number of options K , over all decisions of each run, with the probabilities of the released models. Accuracy, mean confidence (the probability of the top option), and ECE (all in %).
ECE
Log loss
Reduction
Brier
Stable
Accuracy
Most conf.
Model
orig.
avg.
orig.
avg.
(95% CI)
orig.
avg.
(%)
stable
unstable
same cov.
Drex Max
5.8
5.4
0.474
0.460
+0.014 [+0.005, +0.024]
0.237
0.232
92.6
87.2
37.6
87.1
Drex v2.0
3.3
3.4
0.499
0.493
+0.006 [-0.003, +0.016]
0.274
0.271
90.2
84.7
46.8
84.6
Drex v1.5
7.1
6.0
0.579
0.568
+0.011 [-0.001, +0.024]
0.301
0.296
90.8
82.9
37.7
83.3
Jev 1.14
6.9
6.2
0.680
0.636
+0.044 [+0.018, +0.075]
0.300
0.296
93.0
81.8
45.9
82.0
Appendix
Table 20: Order instability and order averaging on the 1,591 decisions of Table 5 . The left columns show the calibration of the original order and of the average of the three orders’ distributions (ECE in %, log loss in nats, Brier), with a bootstrap 95% interval for the reduction in log loss. The right columns show the share of decisions whose top option is the same in all three orders, the accuracy of stable and unstable decisions, and the accuracy of the same number of most confident decisions. The stratified sample over-weights small benchmarks, so its ECE is not comparable with the ECE in Table 4 .
Full suite, Drex Max
80-item sample
Group
Benchmark
Items
Acc.
Skill
Drex Max
Luna Dec.
Perception
CV-Bench
2,038
86.3
77.3
85.0
73.8
BLINK
961
74.1
59.9
80.0
63.7
RealWorldQA
589
84.0
73.9
85.0
75.0
Winoground
800
93.0
86.0
93.8
83.8
Documents
CharXiv
532
86.1
81.5
86.2
78.8
Appendix
Table 21: Decision Index vision suites (public benchmarks). Left: accuracy and chance-corrected skill of Drex Max on all 12,210 items. Right: accuracy of Drex Max and GPT-6 Luna Decisions on the same sample of 80 items per benchmark, with the better of the two in bold. w is the weight of a benchmark in the public vision index; benchmarks without a w have weight 1.
Table 33
Benchmark
Metric
Requests
Drex Max
Drex v2.0
Drex v1.5
Luna Dec.
Knowledge
BBH fixed-option tasks
acc.
5507
93.7
78.3
85.6
66.0
CLadder
acc.
5000
85.7
84.5
86.1
63.5
CRUXEval
acc.
570
85.8
84.6
76.7
66.3
ChessBench
acc.
5000
24.2
23.4
22.0
17.4
GPQA Diamond
acc.
196
56.1
51.5
43.9
48.5
Appendix
Table 24: Decision Index 0.4.1, all public index benchmarks (official metric ×100 , our evaluations). Best per row in bold.
Category
N
Drex Max
Max (text)
Drex v2.0
Drex v1.5
Jev 1.14
d3-turbo
GPT-6 Sol
Opus 5.5
AI agents & evals
154
90.3
90.3
83.1
82.5
90.9
90.9
92.2
96.1
Customer support
67
97.0
97.0
94.0
95.5
97.0
97.0
98.5
98.5
Data & analytics
126
100.0
100.0
97.6
97.6
97.6
99.2
99.2
100.0
Design
30
96.7
26.7
36.7
46.7
33.3
96.7
100.0
100.0
Documents & meetings
87
87.4
87.4
78.2
80.5
85.1
86.2
89.7
88.5
Engineering
150
96.0
96.0
88.7
85.3
97.3
95.3
94.7
97.3
Appendix
Table 25: Decision Bench v4 accuracy by category (%). Drex Max receives the images, whereas Max (text) receives the same textual version of the images as the text-only models.
Decision type
Item that Drex Max is asked about
Meaning of yes
Use of p by the framework
Query routing
Each search query for which the caller lets the framework choose the search strategy.
Answering the query requires a lookup of one specific fact, rather than an investigation that combines facts across many documents or table rows.
If 1−p≥0.80 , the framework passes the query to the search agent; otherwise, it answers the query with fact search.
File relevance
Each file on the shortlist, which holds the 12 files ranked highest by the retrieval system.
The file states the information that the query asks for.
The framework sorts the shortlist by p plus a small bonus for a high retrieval rank.
Table relevance
Each table on the shortlist.
The rows of the table contain the information that the query asks for, or can be used to compute it.
If p is below a fitted threshold, the framework removes the table from the shortlist, unless the table has the highest p on the shortlist.
Passage sufficiency
Each file on the shortlist.
The passages and graph facts that retrieval returned for the file contain the complete answer.
If p is above a high threshold for any file, the framework fetches no files, and the answering model writes the answer from the passages of those files.
Appendix
Table 26: The four types of decisions that Drex Max makes in our document search framework. Each row describes one type of decision. For each item in the second column, the framework asks Drex Max one noul question, and p denotes the probability that Drex Max assigns to the answer yes. The third column states what an answer of yes means, and the fourth column states how the framework uses p .
Figure 13: Inference latency of Drex v1.5. (a) Median latency of the previous and the current version of our inference code. A new question about a cached state of 114–117k tokens is answered faster than an uncached state of 2k tokens can be processed. (b) For an uncached state at the maximum length of 131,072 tokens, almost all of the time is spent in the state pass, which processes the state once per request, and the total exceeds our target of 2 s.
Implementation
Drex DLM
Drex v1.5
Drex v2.0
Drex Max
Python reference server
CUDA, Metal ∗
CUDA
CUDA (Kev)
CUDA
llama.cpp fork
edlm branch
drex-v1.5
drex-v2.0
drex-max-v1 , text only
Ollama fork
nace-edlm branch
drex-v1.5
drex-v2.0
drex-max-v1 , text only
vLLM pooling + in-process head
—
—
supported
—
Hosted API
—
drex-v1.5
drex-v2.0
drex-max
Appendix
Table 27: Validated inference implementations. ∗ Apple silicon requires disabling Metal tensor matmul for the DLM.
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.
We introduce StartLux-Decision, decision models from 0.8B to 27B parameters that combine efficient single-pass prediction with adaptive, confidence-guided reasoning. The Instant models use bidirectional attention in full-attention layers to improve decision accuracy while keeping single-pass inference. The AutoThink models invoke explicit reasoning for uncertain decisions and answer confident cases directly. A three-stage training pipeline establishes shared decision capabilities, specializes the two variants, and equips AutoThink with reasoning through reasoning supervision and reinforcement learning. We introduce Reinforcement Learning for Calibrated Deliberative Decisions (RLCD2), which uses a frozen model and proper scoring rules to reward improvements in decision probability quality and applies a soft penalty to overlong reasoning traces. RLCD2 raises the share of reasoning traces that terminate within the generation limit by up to 48 percentage points and cuts average reasoning length by up to 65%, while largely preserving decision accuracy. On the public subset of Decision Index 0.3, the Instant models outperform the strongest evaluated comparable-scale open-weight decision models by 7.4-16.9 points, and the 27B AutoThink model reaches 73.4. We also introduce DecisionPivot, a programmatically verifiable benchmark of computational decisions built from contrast pairs that differ in a single answer-changing detail. On DecisionPivot, AutoThink outperforms its Instant counterparts by 11.5-18.4 points in pair accuracy and leads among evaluated systems. The models can be run using frameworks such as vLLM and MLX, with quantized versions available to reduce memory usage.Inference code and links to model weights are available at https://github.com/StartLuxLabs/Startlux-Decision .
Xialie Zhuang, Zunhai Su, Jia Pan +3
StartLux · The University of Hong Kong · Southern University of Science and Technology
Large language models make decisions through parallel prefill and sequential decoding. Although effective, this approach incurs decoding costs and does not explicitly expose decision probabilities. Existing task-specific classifiers require per-task training and offer limited generalization to unseen decision schemas. We propose Large Decision Models (LDMs), repurposing language models to return explicit probabilities over labels specified at run time using one prefill forward pass. We first adapt models to this interface through Diverse Decision Supervision (DDS), covering varied domains, rules and labels. To further narrow the performance gap to a reasoning model, Self-Posed Reasoning Distillation (SPRD) adds tasks generated by a frozen teacher, filters them for answer consistency, and trains on label distributions obtained from repeated teacher reasoning. We evaluate on DecisionVista, pairing new-schema tests with held-out human-labelled sources. Across three base models, the recipe improves accuracy over a step-matched baseline by 1.8 and 1.5 percentage points, respectively, achieving a 182-fold speedup in latency over reasoning rollout on Gemma-4-31B. Reasoning remains stronger on the new-schema test, motivating a confidence router for uncertain decisions. On development sets, routing recovers a median 92% of reasoning's accuracy advantage while generating 27-58% of its tokens. Code is available at https://github.com/JianhuiWei7/LDM.
Ziyi Chen, Yan Zhang, Jianhui Wei +2
University of Illinois Urbana-Champaign · National University of Singapore · Zhejiang University +2
Decision needs vary with each application, language, and deployment environment. The Decision Model Factory turns those needs into specialized, sovereign models on demand. A brief recipe guides the training, calibration, verification, and evaluation of a selected backbone, delivering a deployable decision model under the operator's control. The same workflow creates a new specialist whenever the requirements change. Across Japanese decision tasks, compact deployment, and long-context passage ranking, five factory-built models improve upon their backbones; three surpass the strongest results in a fixed public Japanese benchmark comparison. A dedicated runner computes only permitted-answer probabilities and reuses shared context, more than doubling serving speed on matched multi-question workloads with essentially unchanged decision scores. Together, the factory and runner make fast, task-specific decision models practical to build and deploy whenever needed.