Heterogeneous large language models (LLMs) exhibit complementary strengths, motivating routing and multi-model inference methods that seek to identify reliable predictions for each input. However, exploiting this complementarity remains challenging even when a correct prediction is available within the ensemble. We investigate this challenge across heterogeneous LLM ensembles and find that a simple single-round aggregation rule is highly competitive with learned routing methods, yet substantial oracle headroom remains. Further analysis shows that the evaluated single-round response signals provide limited discriminative power for identifying reliable predictions. In contrast, controlled re-presentations of the same question elicit complementary evidence about prediction reliability. Motivated by these findings, we introduce RE-PRESENT, a training-free framework that selectively re-presents disputed questions and directly aggregates candidate-level evidence across alternative views. Across seven multiple-choice benchmarks and two heterogeneous model ensembles, RE-PRESENT improves over strong single-round aggregation while offering a favorable accuracy--cost trade-off compared with self-consistency and multi-model debate.
Figures & tables
Figure 1: Overview of RePresent. The framework first aggregates single-round model evidence, selectively re-presents disputed candidates under alternative views, and combines cross-view evidence for the final prediction.
Method
MedQA
MedMCQA
MMLU
ARC-C
OpenBookQA
MathQA
HeadQA
Avg.
7–9B Ensemble
Average
60.9
57.6
78.1
88.3
85.0
71.7
68.3
72.83
Best Single LLM
66.8
61.5
78.8
91.0
86.0
85.0
75.1
77.74
Majority Vote
66.7
62.5
81.5
92.3
90.4
83.5
74.4
78.76
LLM-Blender
57.6
57.3
78.1
89.1
84.9
74.3
67.1
72.63
GraphRouter
59.0
57.2
79.4
90.0
88.6
66.5
66.1
72.40
Table 1: Per-dataset test accuracy (%) of single-round routing and aggregation methods. Each dataset contains 1,000 test questions, and Avg. denotes the unweighted mean across seven datasets. Average denotes the mean accuracy of the five individual models, while Best Single LLM denotes the highest-performing individual model for each dataset. AgentRouter results are averaged over three seeds. The oracle counts a question as correct whenever at least one ensemble member produces the correct answer. Bold indicates the highest non-oracle accuracy in each column, including ties.
Decision Rule
7–9B
14–32B
Random LLM
40.2
40.5
Best Fixed Model
49.1
45.9
Highest-Confidence LLM
48.5
48.2
LR (Per-Model Signals)
49.8
53.1
LR (All Single-Round Signals)
54.7
52.6
Majority Vote
49.4
51.1
Table 2: Accuracy (%) on disputed questions using single-round evidence. LR uses model-specific or ensemble-level features. Best Fixed Model selects the top-performing model on the training data. Oracle Selection assumes a correct answer whenever any model is correct. Best non-oracle results are bolded; second-best results are underlined.
Reliability Signal
7–9B
14–32B
Single-Round Response Signals
0.722
0.700
Re-presentation Signals
0.708
0.692
Combined
0.745
0.724
Table 3: AUROC for predicting individual model correctness on disputed questions. Combining single-round features with cross-presentation behavior improves discrimination in both ensembles.
Figure 2: Accuracy–token trade-off across two heterogeneous model ensembles. Each point represents average accuracy and total token usage, normalized by the initial five-model inference. Re-Present achieves competitive accuracy with substantially lower token usage than the interactive baselines.
Evidence
7–9B
14–32B
Stage 1 Only
80.31
86.67
+ Reversed View
81.19
87.09
+ Top-Two View
80.91
87.16
+ Both Views
81.51
87.46
Table 4: Average accuracy (%) when incorporating different re-presentation views into Stage 1 evidence.
Decision Rule
7–9B
14–32B
Stage 1 Log-Linear Pooling
54.9
53.1
Select Predicted-Best Model
56.5
53.2
Reliability-Weighted Vote
54.9
51.7
Re-Present Pooling
58.9
57.5
Table 5: Accuracy (%) on disputed questions using different strategies for incorporating re-presentation evidence. Learned reliability strategies use both single-round and re-presentation signals, while Re-Present directly aggregates candidate-level distributions.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Stable
Flip
Ens.
View
n
Acc.
n
Acc.
Gap
7–9B
Top-Two
5692
51.6
4443
25.7
+26.0
Reversed
4926
53.6
5209
27.6
+26.0
14–32B
Top-Two
3424
50.8
2556
26.7
+24.1
Reversed
3067
52.4
2913
28.0
+24.3
Appendix
Table 6: First-pass correctness (%) by cross-presentation stability on disputed model–question pairs. n denotes the number of model–question pairs in each group. Gap denotes the stable–flip accuracy difference in percentage points.
7–9B Ensemble
14–32B Ensemble
Dataset
Gap
AUROC
Gap
AUROC
MedQA
+19.8
0.629
+21.9
0.625
MedMCQA
+8.7
0.558
+17.0
0.609
MMLU
+29.1
0.693
+28.4
0.671
ARC-C
+41.4
0.725
+42.4
0.745
OpenBookQA
+38.5
0.739
+35.4
0.695
Appendix
Table 7: Per-dataset association between reversed-view stability and single-round correctness. Gap denotes stable accuracy minus flip accuracy in percentage points. AUROC evaluates the binary stability indicator as a predictor of single-round correctness.
Model
Gap
AUROC
7–9B Ensemble
Gemma2-9B
+24.3
0.671
Llama3.1-8B
+23.7
0.658
Ministral3-8B
+38.7
0.711
Qwen2.5-7B
+14.0
0.610
Qwen3-8B
+26.1
0.668
Appendix
Table 8: Per-model association between reversed-view stability and single-round correctness on disputed questions. Gap denotes stable accuracy minus flip accuracy in percentage points, while AUROC evaluates stability as a binary correctness indicator.
Confidence
Behavior
7–9B
14–32B
High
Stable
64.3
61.5
High
Flip
36.4
34.1
Low
Stable
53.8
52.8
Low
Flip
24.5
26.6
Appendix
Table 9: Single-round correctness (%) stratified by confidence and joint cross-presentation stability. Stable indicates agreement with the original prediction under both views, while Flip indicates a change under at least one view.
Method
Avg.
Token Usage ( × )
Input
Output
Total
7–9B Ensemble
Selective Debate
81.06
2.77
1.26
1.86
Selective MAD-M 2 (S)
81.53
7.44
3.92
5.32
Selective MAD-M 2 (O)
79.93
2.80
1.56
2.05
Selective Self-Consistency
80.79
1.63
1.34
1.46
Appendix
Table 10: Average accuracy (%) and token usage under the same disagreement-based selective trigger. Token usage includes the initial inference and is normalized by the initial five-model pass ( 1.00× ). Input, Output, and Total denote input, generated, and combined token volumes, respectively. Bold and underline indicate the best and second-best values within each ensemble, respectively.
Ensemble
Condition
Calls
Extra Tok.
Acc.
Δ
Stab. AUROC
7–9B
Same presentation, greedy
1
23%
80.74
+0.37
0.614
Self-consistency, T=0.7
1
23%
80.31
-0.05
0.624
Reversed
1
23%
81.19
+0.81
0.656
Top-Two
1
20%
80.91
+0.53
0.669
Self-consistency, two samples
2
46%
80.47
+0.10
0.647
Top-Two + Reversed
2
43%
81.51
+1.15
0.697
Appendix
Table 11: Controls comparing additional inference under the original presentation with alternative candidate-space views. Calls denotes additional inference calls per model on disputed questions. Extra Tok. reports additional combined input and generated tokens relative to the initial five-model inference. Acc. is overall average accuracy (%), and Δ denotes the improvement over Stage 1. Stab. AUROC evaluates cross-presentation stability as a correctness indicator.
Ensemble
View
Flip Rate
Stab. AUROC
Acc.
Δ
7–9B
Reversed
51.4
0.656
81.19
+0.81
Random 1
47.3
0.649
81.39
+1.01
Random 2
48.1
0.642
80.77
+0.40
Random 3
46.9
0.642
81.41
+1.05
Cyclic Shift
47.7
0.646
81.14
+0.77
Top-Two
43.8
0.669
80.91
+0.55
Appendix
Table 12: Alternative presentation controls on disputed questions. Flip Rate denotes the percentage of model–question pairs whose prediction changes relative to Stage 1. Stab. AUROC evaluates stability as a binary correctness indicator. Acc. denotes overall average accuracy (%), and Δ is the improvement over Stage 1 in percentage points.
Ensemble
Pattern
n
Stage 1
RePresent
Δ
W → C
C → W
7–9B
5–0
3338
95.5
95.6
+0.1
0.1
0.0
4–1
1635
81.0
80.9
-0.1
1.2
1.3
3–2
811
57.1
63.5
+6.4
13.1
6.7
3–1–1
632
61.1
62.0
+0.9
5.7
4.7
2–2–1
434
45.6
50.5
+4.8
13.1
8.3
2–1–1–1
146
44.5
45.2
+0.7
9.6
8.9
Appendix
Table 13: Accuracy (%) and prediction transitions by single-round voting pattern. Stage 1 and RePresent denote accuracy before and after re-presentation, respectively. Δ is the accuracy difference in percentage points. W → C and C → W denote the percentages of questions corrected and made incorrect, respectively, within each pattern.
Large Language Models (LLMs) reason in free-form chain-of-thought (CoT) without revealing which steps were well supported and which alternatives were rejected. In high-stakes domains, users then cannot judge if an answer rests on sound reasoning. Majority voting and multi-agent debate improve answers using model agreement, but do not expose agreement on individual steps. Reasoning Consensus (RC) extracts a typed reasoning graph from each CoT, merges equivalent steps across graphs from diverse models, and weights every step by how many chains attest to it. RC reveals this step-wise support and the varied justifications behind each answer. On 7 benchmarks across law, science, medicine, narrative reasoning and logic, RC matches majority voting and debate in accuracy (69.4% vs. 69.0-70.3%). An LLM judge prefers RC's best-supported graph to other graphs for the same answer. In a user study, participants read RC graphs faster than debate transcripts or vote tallies, and 14 of 18 preferred them for future use.
LLM ensemble gains are often measured against a single baseline, but the value attributed to additional candidates can change with the comparison used. We study an LLM aggregator under four matched conditions: Direct (no candidates), Single-slot (one candidate in a multi-slot interface), Duplicate (one candidate repeated), and Multi-candidate (separately generated candidates). Under either non-diverse control, overall candidate gain (CIV) splits algebraically into an additional-candidate contrast (DSV) and a candidate-context contrast (CCV); the split, but not CIV, depends on the chosen control. On MultiHopRAG model diversity (k=2), the repetition gap Rep=-8.33 pp [-13.25, -3.75] quantifies the difference between gains measured against Duplicate and Single-slot. On a matched FRAMES stronger-panel extension (n=200, k=3, aggregator output limit 4096), paired Delta Rep=+9.67 pp [+6.00, +13.50]; within panels, Rep changes sign (original -6.33; stronger +3.33). In FRAMES evidence aggregation, candidate-context contrasts remain strongly negative under both non-diverse controls. Our results show why ensemble evaluations should state the reference control before attributing gains to additional candidates.
Selecting the correct answer from a pool of candidate reasoning chains is the engine of test-time scaling, yet the standard selectors each carry a cost: self-consistency inherits the errors of the single model it resamples, and trained reward models need labeled data and transfer poorly off-distribution. We study a third signal, free at inference time: cross-model consensus, the degree to which independently trained models, each solving the problem once, agree on a final answer. We treat the panel as an LLM-jury, in which the structure of agreement, not any model's score of another, is the verification signal. Across seven benchmarks it selects correct answers better than self-consistency and far better than a model scoring its own candidates: on competition math it closes the entire gap to an oracle selector, while self-scoring closes almost none. The mechanism is error decorrelation: independently trained models err differently, so their wrong answers scatter while the correct one accumulates agreement. We make this precise with a parameter-free law, derived in closed form, that predicts consensus accuracy from three measured panel statistics to a mean absolute error of 0.03 and exposes the method's ceiling: a shared-error floor where models share a misconception, near zero on math but non-trivial on science. Against four trained verifiers spanning discriminative, outcome, and generative reward models, the free LLM-jury matches the strongest inside their math training domain and is the top selector outside it. Cross-model consensus is thus a verifier we can characterize in advance: a law that says when to trust it, and a floor that marks where it cannot.