Organizations: University of Wisconsin–Madison · School of Artificial Intelligence, Jilin University · School of Data Science, Chinese University of Hong Kong, Shenzhen
Typed decision models are language models that answer a question with a fixed set of possible answers by returning a probability for each answer instead of generating text. We study them on tabular data and compare four decision models with three large language models (LLMs) on row classification, fact verification and anomaly detection. Every request is sent in its original form and under 21 types of rewrite that keep its content, such as renaming the options or writing the table in another format. The decision models perform as well as the LLMs at a lower cost. Three of them are also far more robust to how the question is formulated. No model we test is robust to how the table is organized. For every model, at least a third of the row decisions change under some rewrite of the table or the question. The decisions to which a decision model gives low probability are the ones most likely to change. Decision models are a practical choice for tabular decisions, but their robustness remains an open problem.
Figures & tables
Decision models
Research question
Jev, Clef, Kev-27B
GPT-6 Luna Decisions
LLMs
RQ1 Are decision models practical? (§ 3 )
✓ As accurate as LLMs and best on two of three tasks, at far lower cost
✓ As accurate as GPT-6 Luna and cheaper on rows and samples
✓ Best on row classification, but a decoding pass per decision
RQ2 Are they robust to the formulation of the question? (§ 4 )
✓ Option names and order barely matter
✗ Reversing the options changes up to 23%
✗ Option names and order change up to 30%
RQ3 Are they robust to how the table is organized? (§ 5 )
✗ Format and row and column order change many decisions
✗ Likewise
✗ Likewise
RQ4 Does their confidence indicate which decisions are sensitive to rewrites? (§ 6 )
✓ Low confidence marks the decisions a rewrite changes
Partly. As well as an LLM’s best readout
Partly. Only with ten draws or token probabilities
Table 1: Our four research questions and their answers for the decision models and the LLMs.
Task
Input
Decision
Data
Row classification
One sample and a description of its features
Its class
10 datasets, 3,481 rows
Fact verification
A table and a claim about it
Supported or refuted
TabFact, 2,000 claims
Anomaly detection
One sample, a description of the data and statistics of normal samples
Normal or anomalous
ReTabAD, 20 sources, 5,059 samples
Table 2: The three tasks: row classification on the ten datasets of TabLLM and FeatLLM ( Hegselmann et al., 2023 ; Han et al., 2024 ) , fact verification on TabFact ( Chen et al., 2020 ) and anomaly detection on ReTabAD ( Yoon et al., 2026 ) . Metrics: Macro-F1, accuracy and AUROC, averaged over datasets or sources.
Question , options with their descriptions
original
{no, yes} with descriptions
options reversed
{yes, no}
renamed A, B
{A, B}, descriptions kept
synonyms
{false, true}
yes/no question
"Is the target class yes?"
paraphrased
"Determine the class …"
Figure 1: One row-classification request (blood donation) under the rewrites of Figure 3 , shortened to the part that each rewrite changes. Shuffling or reversing the columns permutes their names, descriptions and values together.
Figure 2: In-context examples (RQ1). Grey lines show the best of eight classifiers or five anomaly detectors trained on the same rows, and dotted lines the best trained on all training rows. max stands for as many rows as fit in the context. In (c), stats is the original request of ReTabAD, which summarizes the normal training samples by statistics. none leaves these statistics out, and +16 and +max add raw normal samples.
Figure 3: Row classification, per rewrite: decisions that differ from the model’s original decision (left) and the change of Macro-F1 (right). The upper block shows RQ2 and the lower block RQ3. Filled markers are decision models and hollow markers LLMs. Clef’s point for option order is hollow because Clef sorts the options.
performance
generated tokens
Model
Cl.
TF
AD
Cl.
TF
AD
Jev
52.1
89.7
74.6
none: one pass b
Clef
58.4
85.0
74.7
none: one pass
Kev-27B
55.2
88.8
81.2
none: one pass
GPT-6 Luna Decisions
61.2
78.7
80.5
none c
GPT-6 Luna
62.5
79.0
77.3
5.6
11.8
87.5
Table 3: RQ1: performance under the original request and the tokens each model generates per decision (an anomaly batch carries 15 decisions). Cl. = row classification (Macro-F1), TF = TabFact (accuracy), AD = anomaly detection (AUROC). a Best of eight classifiers trained on the whole training split, LightGBM (Cl.), and of five anomaly detectors, k -nearest neighbours (AD, 19 of 20 sources). We train no model for TabFact. b As stated by its provider. Jev’s architecture is not public. c No output tokens are billed.
seconds
US$ / million
Model
Cl.
TF
AD
Cl.
TF
AD
Jev
0.08
1.33
0.09
10
38
16
GPT-6 Luna Decisions
0.10
1.60
0.11
24
61
37
GPT-6 Luna
1.62
1.59
0.73
81
62
68
Table 4: Latency and price per decision of the closed models: mean wall-clock seconds, and US$ per million decisions as billed, on the original requests. A Jev or GPT-6 Luna Decisions request answers 16 rows or a batch of 15 samples, and a GPT-6 Luna request one row, one claim or one batch.
Model
Row classif.
TabFact
Anomaly det.
Jev
33.8
19.3
21.1
Clef
31.7
23.0
25.8
Kev-27B
32.2
18.3
17.7
GPT-6 Luna Decisions
39.6
38.3
29.9
GPT-6 Luna
42.8
60.8
63.4
Qwen3.8-27B
31.8
27.4
23.3
Table 5: Items (%) whose decision changes under at least one information-preserving rewrite of the table (6, 10 and 8 types of rewrite, Table 12 ). GPT-6 Luna is run at temperature 1, so its shares include its re-send rate.
Model
Confidence
Cl.
TF
AD
Trained from the same backbone
Clef
probability
1.4
4.1
0.7
Kev-27B
probability
3.1
2.5
0.1
Qwen3.8-27B
ten draws
2.6
9.7
4.8
stated
23.4
29.2
9.2
The same model through two interfaces
Table 6: Of each model’s 25% most confident items, the share (%) changed by at least one rewrite of the table. Cl. = row classification, TF = TabFact, AD = anomaly detection.
Figure 4: Re-querying the least confident rows (row classification). For the x% of rows with the lowest confidence, the answer in the original format is replaced by the combination of the answers in the original and the four other formats. The y-axis gives the share of the Macro-F1 gain obtained when every row is re-queried, and the legend gives that gain in points.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Decision models: observed flip rate against the rate predicted when a rewrite adds item-independent noise to the margin, per margin decile and task.
predicts a flip
predicts an error
Model
Row classif.
Anomaly
TabFact
Row classif.
Anomaly
TabFact
Jev
0.87 [0.86, 0.88]
0.89 [0.89, 0.90]
0.89 [0.88, 0.91]
0.59
0.65
0.82
Clef
0.85 [0.85, 0.86]
0.89 [0.88, 0.89]
0.87 [0.86, 0.89]
0.67
0.61
0.84
Kev-27B
0.85 [0.84, 0.86]
0.92 [0.92, 0.93]
0.90 [0.88, 0.91]
0.70
0.70
0.84
GPT-6 Luna Decisions
0.82 [0.81, 0.83]
0.83 [0.83, 0.84]
0.81 [0.80, 0.83]
0.74
0.70
0.78
GPT-6 Luna
0.83 [0.82, 0.84]
0.71 [0.70, 0.72]
0.73 [0.72, 0.75]
0.68
0.62
0.69
Appendix
Table 9: How well a model’s confidence in its original decision predicts (left) that an information-preserving rewrite changes the decision and (right) that the original decision is wrong (flip AUROC and error AUROC; 0.5 is chance). Left: pooled over the rewrites of Figure 3 and the table formats, with 95% cluster-bootstrap intervals. Confidence: for the decision models, the returned probability of the chosen option; for the LLMs, the share of ten draws at temperature 1 that give the original decision. † Strict parse: an answer such as “Yes.” to the yes/no question counts as invalid, hence as changed. Ignoring case and a final full stop: Llama-3.3-70B on row classif. 0.66 [0.65, 0.67].
Model
Ten draws
Stated
Answer prob.
Row classification
GPT-6 Luna
0.83
0.72
0.71
Qwen3.8-27B
0.77
0.60
0.65
Llama-3.3-70B
0.64
0.50
0.73
weakest of Jev, Clef and Kev-27B: Kev-27B, 0.85
GPT-6 Luna Decisions: 0.82
Appendix
Table 10: Every readout of an LLM’s confidence, on the same rewrites: flip AUROC (rewrites shared by both kinds of model). Ten draws : share of ten draws at temperature 1 that give the original decision, the readout of the main text. Stated : the confidence the model gives when asked, 0–100 (anomaly detection: the confidence in its answer). Answer prob. : probability of the first token of the generated label, or the stated anomaly score’s distance from 0.5; for GPT-6 Luna, agreement of three draws. Bold: the best readout of each model and task. –: not available.
arXiv
Table input (application or setting)
2609.23136
telemetry tables (6G network intents)
2609.24395
video feature records (video quality)
2609.27535
survey micro-data as personas (social science)
2609.33689
KPM records (5G network control)
2609.36154
accelerometer features (activity recognition)
2609.37470
joined tables or nested records (synthetic world)
Appendix
Table 11: Of the 112 arXiv papers on decision models posted between 15 September and 8 October 2026, the fourteen whose main input is a table or table rows.
Task
Code
Rewrite
Source
Row classification
COLS
Columns shuffled: one fixed permutation per dataset; field descriptions, names and values permuted together
order
COLR
Columns in reverse order
order
LIST
One line per field, “- name: value” (TabLLM list)
format
TEXT
One sentence per field, “The name is value.” (TabLLM text)
format
JSONOBJ
JSON object {name: value}
format
MD
One-row Markdown table (header and value row)
format
Appendix
Table 12: Rewrites of the table that keep the information needed for the decision. The original format is, for row classification, field descriptions followed by one array of values; for TabFact, the official format with a header line and # between cells; for anomaly detection, the serializer of the ReTabAD authors with 15 samples per batch. Source: order of rows and columns, or format of the table. a Adds a decimal of precision, so it is left out of Tables 5 and 6 . b Can change the truth of claims that refer to positions ( first , last , …); such claims are left out of Tables 5 and 6 .
decision models
LLMs
Jev
Clef
Kev-27B
GPT-6 Luna Decisions
GPT-6 Luna
Qwen3.8-27B
Llama-3.3-70B
Condition
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Row classification (Macro-F1)
original request
–
52.1
–
58.4
–
55.2
–
61.2
–
62.5
–
52.2
–
44.0
identical re-send
4.7
52.4
1.6
58.0
0.6
54.5
0.0
61.2
13.3
63.0
1.1
52.9
0.4
44.1
options in reverse order
5.9
52.3
1.6
58.0
5.9
53.9
22.7
60.1
13.7
62.5
13.6
54.1
26.8
51.4
Appendix
Table 13: How the question is asked, raw results. Flip : share of items whose decision differs from the model’s own decision under the original request (%; invalid outputs count as flips). Score : the task metric × 100 (row classification: Macro-F1 averaged over the 10 datasets; TabFact: accuracy; anomaly detection: AUROC averaged over the 20 sources). Identical re-send : the original request sent again, which gives the re-send rate. Conditions below the line remove or mislead information. Clef sorts the options, so reversing them cannot reach it. On row classification the LLMs receive no option descriptions, and on anomaly detection no options. Llama-3.3-70B answers the yes/no question mostly with “Yes” or “Yes.”, which the strict parse counts as invalid; with a parse that ignores case and a final full stop its flip rate there is 13.0%. Invalid outputs above 5% (%): Llama-3.3-70B on row classification, yes/no question: 54.8; GPT-6 Luna on TabFact, option names A, B, …: 9.8; GPT-6 Luna on TabFact, option names replaced by synonyms: 43.7; GPT-6 Luna on TabFact, no task description: 6.6; GPT-6 Luna on ReTabAD, paraphrased instructions: 7.0; Llama-3.3-70B on ReTabAD, paraphrased instructions: 9.5; GPT-6 Luna on ReTabAD, no task description: 5.2. –: does not apply.
decision models
LLMs
Jev
Clef
Kev-27B
GPT-6 Luna Decisions
GPT-6 Luna
Qwen3.8-27B
Llama-3.3-70B
Condition
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Row classification (Macro-F1)
original request
–
52.1
–
58.4
–
55.2
–
61.2
–
62.5
–
52.2
–
44.0
identical re-send
4.7
52.4
1.6
58.0
0.6
54.5
0.0
61.2
13.3
63.0
1.1
52.9
0.4
44.1
columns shuffled
15.9
51.6
15.0
57.8
15.6
54.8
18.5
57.0
20.1
60.3
20.4
52.5
17.2
46.3
Appendix
Table 14: Order of rows and columns, raw results. Flip : share of items whose decision differs from the model’s own decision under the original request (%; invalid outputs count as flips). Score : the task metric × 100 (row classification: Macro-F1 averaged over the 10 datasets; TabFact: accuracy; anomaly detection: AUROC averaged over the 20 sources). Identical re-send : the original request sent again, which gives the re-send rate. Shuffling the questions of a request applies only to the decision models, which answer 16 questions per request; the LLMs receive one row per prompt. Reordering rows can change the truth of TabFact claims with position words. Invalid outputs above 5% (%): GPT-6 Luna on TabFact, columns shuffled: 6.7; GPT-6 Luna on TabFact, columns reversed: 7.2; GPT-6 Luna on TabFact, rows shuffled: 5.9; GPT-6 Luna on TabFact, rows reversed: 6.8; GPT-6 Luna on TabFact, table transposed: 8.2; GPT-6 Luna on ReTabAD, columns shuffled: 7.3; GPT-6 Luna on ReTabAD, columns reversed: 7.1; GPT-6 Luna on ReTabAD, rows shuffled: 8.8; GPT-6 Luna on ReTabAD, rows reversed: 8.0. –: does not apply.
Figure 6: Row classification, by rewrite. Left: flip AUROC. Middle: among flips that change correctness, the share that turn a right answer wrong. Right: standard deviation of the change of the decision models’ margin. Clef is omitted for option order, since it sorts the options.
decision models
LLMs
Jev
Clef
Kev-27B
GPT-6 Luna Decisions
GPT-6 Luna
Qwen3.8-27B
Llama-3.3-70B
Condition
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Flip
Score
Row classification (Macro-F1)
original request
–
52.1
–
58.4
–
55.2
–
61.2
–
62.5
–
52.2
–
44.0
identical re-send
4.7
52.4
1.6
58.0
0.6
54.5
0.0
61.2
13.3
63.0
1.1
52.9
0.4
44.1
list (“- name: value”)
15.0
56.0
14.0
60.3
16.1
54.6
17.2
64.7
17.7
66.5
17.9
58.4
21.3
47.9
Appendix
Table 15: Format of the table, raw results. Flip : share of items whose decision differs from the model’s own decision under the original request (%; invalid outputs count as flips). Score : the task metric × 100 (row classification: Macro-F1 averaged over the 10 datasets; TabFact: accuracy; anomaly detection: AUROC averaged over the 20 sources). Identical re-send : the original request sent again, which gives the re-send rate. Conditions below the line remove or mislead information. Rendering numbers with three decimals adds precision. Invalid outputs above 5% (%): GPT-6 Luna on TabFact, CSV: 5.1; GPT-6 Luna on TabFact, JSON records: 7.3; GPT-6 Luna on TabFact, TAPEX linearization: 8.9; GPT-6 Luna on ReTabAD, JSON object: 9.7; GPT-6 Luna on ReTabAD, list (“- name: value”): 7.7; GPT-6 Luna on ReTabAD, text (“The name is value.”): 9.2; GPT-6 Luna on ReTabAD, Markdown table: 5.6; GPT-6 Luna on ReTabAD, numbers with three decimals: 8.2; GPT-6 Luna on ReTabAD, values without names: 10.9. –: does not apply.
Jev
Clef
Kev-27B
GPT-6 Luna Decisions
trained on the same examples
Row classification (Macro-F1)
Score
Δ
Score
Δ
Score
Δ
Score
Δ
LogReg
LightGBM
Best
TabPFN
no examples (zero-shot)
52.1
–
58.4
–
55.1
–
61.2
–
–
–
–
–
4 examples
60.2
+8.1
60.0
+1.6
59.0
+3.8
59.6
− 1.6
51.8
45.3
55.2
51.5
16 examples (the original ones)
60.0
+7.9
61.7
+3.2
62.7
+7.6
60.4
− 0.8
58.6
51.8
58.7
59.5
64 examples b
65.3
+12.2
68.2
+9.0
69.3
+12.2
65.4
+3.3
69.2
69.1
69.4
70.8
as many as fit (30–681)
66.7
+14.6
70.6
+12.1
73.0
+17.6
69.9
+8.7
72.2
76.0
76.3
77.8
Appendix
Table 16: In-context examples, raw results. Examples are labelled training rows placed in the request (for TabFact, other tables with a claim and its label; for anomaly detection, normal training records after the normal-sample statistics); the sets are nested. Score : task metric × 100. Δ : change of the score from no examples (anomaly detection: from the original request), on the same rows. The other methods are trained on exactly the same examples: logistic regression with cross-validated regularization, LightGBM with few-row settings, and the best of eight trained classifiers (logistic regression, default and tuned; random forest; k -nearest neighbours; histogram gradient boosting, LightGBM, CatBoost and XGBoost), the line of Figure 2 ; on the normal samples, isolation forest, one-class SVM, k -nearest neighbours and the best of five detectors. TabPFN : TabPFN-3.5, a tabular foundation model that reads the same examples in its context without training (not run on the whole split); TabPFN v2, run the same way, scores 52.3, 57.3, 69.3, 76.4 with 4, 16, 64 and as many examples as fit. b Without communities (about 100 features: 64 examples exceed the budget); the trained methods are averaged over the same nine datasets. The trained anomaly methods cover 19 of the 20 sources (covertype left out).
GPT-6 Luna
Qwen3.8-27B
Llama-3.3-70B
Row classification (Macro-F1)
Score
Δ
Score
Δ
Score
Δ
no examples (zero-shot)
62.5
–
52.2
–
44.0
–
4 examples
64.9
+2.4
59.2
+7.0
57.6
+13.5
16 examples (the original ones)
64.5
+2.0
59.5
+7.3
56.7
+12.6
64 examples b
66.2
+3.0
65.4
+10.4
57.4
+12.5
as many as fit (30–681)
69.3
+6.8
67.1
+14.9
60.7
+16.6
Appendix
Table 17: In-context examples, raw results of the LLMs; setup, Score and Δ as in Table 16 , where the methods trained on the same examples are listed. b Without communities (about 100 features: 64 examples exceed the budget).
Figure 7: Left: flip AUROC within each dataset or source; bars are medians. Right: for each row classification dataset, the mean confidence of the original decisions against the share of decisions changed by the information-preserving rewrites.
confidence band (1: lowest 25%)
noise
Task
Model
Items
k
All
1
2
3
4
Re-send
Bound
Question
Row classif.
Jev
3,481
6
33.8 [32.2,35.4]
78.0 [75.3,80.7]
41.6
14.7
0.9 [0.3,1.5]
4.7
25.3
20.2 [18.9,21.6]
Clef
3,481
6
31.7 [30.1,33.2]
77.5 [74.8,79.9]
37.2
10.7
1.4 [0.7,2.3]
1.6
9.0
17.3 [16.2,18.5]
Kev-27B
3,481
6
32.2 [30.6,33.7]
76.7 [74.2,79.6]
40.0
9.0
3.1 [2.0,4.3]
0.6
3.6
17.1 [15.8,18.2]
GPT-6 Luna Decisions
3,481
6
39.6 [38.2,41.2]
80.9 [78.3,83.6]
52.3
23.2
1.8 [1.0,2.8]
0.0
0.0
36.7 [35.1,38.2]
GPT-6 Luna a
3,481
6
42.8 [41.2,44.3]
88.7 [86.6,90.9]
41.5
20.6
20.2 [17.6,22.7]
13.3
57.5
34.9 [33.3,36.4]
Appendix
Table 18: Items whose decision depends on how the table is organized (%). All : share of items whose decision changes under at least one of the k information-preserving table rewrites of Table 12 ; 1–4 : the same for the items sorted by the model’s confidence in its original decision and split into four bands of 25%, 1 the lowest (ties broken at random with a fixed seed). Re-send : share of decisions changed by an identical re-send of the original request ( p ). Bound : 1−(1−p)k , the share that would change at least once if every rewrite flipped independently at rate p ; an upper bound on what noise alone explains, since re-send flips concentrate on the same low-margin items. Question : share of items changed by at least one rewrite of the question (options reordered or renamed, yes/no question, option descriptions removed, paraphrased instructions). Brackets (shown for all items, the outer bands and the question): 95% cluster-bootstrap intervals (Poisson(1) weights, 500 replicates; clusters are classification rows, anomaly-detection samples and TabFact tables). Invalid outputs count as changes. Numbers with three decimals and TabFact claims with position words are left out. Jev’s original TabFact decisions carry no probabilities. Confidence: for the decision models, the returned probability of the chosen option; for the LLMs, the share of ten draws at temperature 1 that give the original decision. On anomaly detection the LLMs receive no options, so only one question rewrite (paraphrase) applies to them. a Its decisions are themselves draws at temperature 1, so its counts are dominated by temperature noise. b Mostly answers such as “Yes.” to the yes/no question, which the strict parse counts as invalid.
Figure 8: Items whose decision changes under at least one information-preserving rewrite of the table. Items are sorted by the model’s confidence in its original decision and split into four bands of 25%; 95% cluster-bootstrap intervals.
Figure 9: Share of decisions changed by a rewrite, for items ranked by the model’s confidence in its original decision and split into bands of 10% (information-preserving rewrites pooled). Confidence as in the main text: the decision models’ returned probability and the LLMs’ share of ten draws that give the original decision; tied values are split at random.
Model
Right (%), top / low 25%
stable, right
stable, wrong
fragile, right
fragile, wrong
Row classification
Jev
67 / 48
44 / 0.79
23 / 0.78
16 / 0.59
18 / 0.56
Clef
86 / 52
52 / 0.84
16 / 0.78
16 / 0.61
16 / 0.59
Kev-27B
85 / 43
48 / 0.76
19 / 0.67
14 / 0.55
19 / 0.54
GPT-6 Luna Decisions
94 / 49
49 / 0.91
11 / 0.82
21 / 0.68
19 / 0.62
GPT-6 Luna
82 / 46
48 / 0.98
9 / 0.95
23 / 0.77
20 / 0.65
Appendix
Table 19: Confidence against correctness and fragility (RQ4). Second column: share of right decisions among the 25% of decisions with the highest and the lowest confidence. Other columns: share of decisions (%) and their mean confidence in each group; a decision is fragile when at least one rewrite of the table flips it. Confidence as in § 6 .
Every major data modality now has a foundation model that understands it natively: text has language models, images have vision models, audio has audio models. Tabular data, the modality on which many consequential real-world AI decisions are made, does not. Every approach to tabular AI today, from gradient-boosted trees to the latest tabular foundation models, requires a preprocessing pipeline before any model can consume the data. None of them understand tabular data as a modality. We introduce the Data Language Model (DLM), the missing foundation model for tabular data. A DLM understands tables the way a language model understands sentences: natively, without serialization or preprocessing, directly from raw cell values. It is the tabular data layer on which AI models, agents, and vertical AI applications can be built, eliminating the preprocessing pipelines that currently stand between raw data and every AI system that consumes it. We present Schema-1, the first DLM: a 140M parameter model trained on more than 2.3M synthetic and real-world tabular datasets. Schema-1 outperforms gradient-boosted ensembles, AutoML stacks, and the tabular foundation models we evaluate on established row-level prediction benchmarks. On missing value reconstruction it achieves lower reconstruction error than all classical statistical methods and frontier large language models on mean performance across conditions, establishing that structural understanding of a dataset's own distributional geometry is more useful for imputation than world knowledge encoded in language. It identifies the industry sector of any unseen dataset from raw cell values alone, reliably across any domain, a task no prior tabular model can perform. It is the native tabular understanding layer that has been missing from the AI stack.
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
Typed decision models (TDMs) return probability distributions over caller-defined options without generating text. TypeSafe released Jev, a commercial typed decision model, on 15 September 2026, and a small body of evaluation and replication work appeared within days. We review 28 papers posted between 19 and 24 September and relate their findings to earlier work on label-probability classification, constrained decoding, reranking, calibration, and model cascades. In this early literature, the typed readout itself has not shown an independent accuracy advantage over comparable label-probability readouts. Jev's clearest gains are in latency and cost, while accuracy gaps remain on harder tasks. In practical deployments, confidence is often used to decide when to defer to a stronger model or a human. We use recurring weaknesses in these studies to derive a 14-item evaluation checklist for future TDM work. Because the evidence covers only the first nine days after the release of one hosted model, the review should be read as an early evidence map rather than a settled assessment of the model class.