Pass@k and other methods of scaling inference compute can improve language model performance in domains with external verifiers, including mathematics and code, where incorrect candidates can be filtered reliably. This raises a natural question: can we similarly scale compute to elicit gains in truthfulness for domains without convenient verification? We show that across five benchmarks and models, surprisingly, it cannot. Even at 25x the inference cost of naive sampling, polling-style aggregation yields no consistent accuracy gains over single-sample baselines and often amplifies shared misconceptions. We find that under uncertainty, models are better at predicting what other models will say within model ensembles than at identifying what is true, revealing a separation between social prediction and truth verification. Across models and benchmarks, aggregation fails to provide a robust truth signal because language model errors are strongly correlated. The source of correlation goes beyond any individual benchmark: we show that even when conditioned on out of distribution random strings and asked to produce pseudo-random outputs, different models produce correlated outputs. Confidence-based weighting provides no benefit because self-reported confidence fails to reliably distinguish correct from incorrect answers. These results delineate a boundary for inference-time scaling: in verified domains, additional samples provide more candidates for a verifier to filter; in unverified domains, additional samples merely reinforce shared misconceptions.
Figures & tables
Figure 1: No ensemble aggregation method consistently outperforms majority voting across benchmarks. We compare five aggregation rules across four benchmarks using five-model ensembles (125 votes per question). Some methods improve performance on individual tasks, but none dominates overall. Error bars show 95% bootstrap confidence intervals.
Figure 2: Models agree with each other more reliably than they agree with truth. Each panel reports Pearson correlations between binary answers from different model families and the ground-truth label (“Truth”), computed over questions in the benchmark. Across datasets, inter-model correlations are consistently positive, but correlation with Truth varies sharply by task, showing that agreement is not a stable proxy for correctness.
Figure 3: Models exhibit correlated behavior even when no ground truth exists. Cohen’s κ ( Cohen, 1960 ) between pairs of models on random strings with forced-choice answers shows stable above-chance agreement, with a similar correlation structure across temperatures, indicating shared inductive biases rather than shared knowledge.
Figure 4: Self-reported confidence is poorly aligned with correctness. Reliability diagrams across four verifier-absent benchmarks. Confidence increases faster than accuracy and does not reliably distinguish correct from incorrect answers.
Figure 5: Agreement remains high under low confidence. Mean confidence for the most popular answer versus its agreement fraction. Deviations from y=x show that consensus persists even as confidence decreases.
Figure 6: Predicted popularity tracks agreement, not truth. Predicted vote share for the correct option (x-axis) versus observed agreement (y-axis). Points denote models; error bars show 95% bootstrap confidence intervals. The dashed line indicates perfect calibration ( y=x ). Models accurately predict consensus even when consensus does not reliably indicate correctness.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Split
Filter/Mapping
Answer
#Q
HLE
test
exactMatch & yes/no only
YES/NO
35
BoolQ
validation
bool → true/false
TRUE/FALSE
100
Com2Sense
validation
label → true/false
TRUE/FALSE
100
Predict-the-Future
train
normalize yes/no
YES/NO
100
Appendix
Table 1: Benchmark construction and filtering used in experiments.
Model
Individual Avg.
Direct Majority
Highest Conf
Conf Weighted
Pred Weighted
Surp. Popular
HLE (T=1.0)
Gemma-3-4B
27.3% [25.3, 29.4]
28.7% [14.3, 42.9]
25.8% [14.3, 40.1]
28.6% [14.3, 42.9]
37.4% [20.0, 51.5]
28.7% [14.3, 45.7]
GPT-OSS-20B
10.1% [8.3, 11.9]
11.7% [2.9, 22.9]
12.0% [3.0, 24.2]
9.2% [0.0, 21.2]
8.5% [0.0, 17.2]
11.2% [2.9, 22.9]
GPT-OSS-120B
11.7% [10.2, 13.3]
8.7% [0.0, 17.2]
2.8% [0.0, 8.6]
8.4% [0.0, 20.0]
5.6% [0.0, 14.3]
8.4% [0.0, 20.0]
Qwen3-32B
28.2% [26.2, 30.2]
25.1% [11.4, 40.0]
22.9% [8.6, 37.1]
17.3% [5.7, 28.6]
19.6% [5.7, 34.3]
22.8% [8.6, 37.1]
Qwen3-235B
21.4% [19.5, 23.4]
20.2% [8.6, 34.3]
14.6% [2.9, 28.6]
14.6% [2.9, 25.7]
22.8% [8.6, 37.1]
25.4% [11.4, 40.0]
Appendix
Table 2: Per-model results (T=1.0) on HLE and BoolQ. Values are accuracy with 95% bootstrap CIs.
Model
Individual Avg.
Direct Majority
Highest Conf
Conf Weighted
Pred Weighted
Surp. Popular
Predict-the-Future (T=1.0)
Gemma-3-4B
52.6% [51.3, 53.8]
53.8% [44.0, 63.0]
53.0% [43.0, 63.0]
52.0% [43.0, 62.0]
53.1% [43.0, 63.0]
52.1% [43.0, 62.0]
GPT-OSS-20B
50.1% [48.5, 51.6]
49.3% [39.0, 59.0]
51.4% [42.4, 61.6]
51.8% [42.4, 61.6]
50.7% [41.0, 60.0]
47.7% [37.0, 58.0]
GPT-OSS-120B
49.3% [47.1, 51.3]
45.4% [34.7, 54.8]
54.0% [44.8, 63.5]
46.9% [36.5, 56.2]
51.8% [42.0, 61.0]
50.4% [41.1, 61.1]
Qwen3-32B
49.8% [48.4, 51.1]
50.0% [40.0, 60.0]
48.8% [39.0, 58.0]
48.1% [38.0, 58.0]
41.0% [32.0, 50.0]
50.1% [40.0, 60.0]
Qwen3-235B
51.8% [50.4, 53.1]
52.1% [42.0, 61.0]
55.2% [45.0, 65.0]
53.8% [44.0, 63.0]
48.9% [39.0, 59.0]
52.0% [42.0, 62.0]
Appendix
Table 3: Per-model results (T=1.0) on Predict-the-Future and Com2Sense. Values are accuracy with 95% bootstrap CIs.
Figure 7: Per-model aggregation accuracy on factual and expert benchmarks. Accuracy by aggregation rule at temperatures T∈{0.7,1.0} . Bars correspond to individual-sample average, direct majority vote, highest-confidence selection, confidence-weighted vote, predicted-popularity weighting, and Surprisingly Popular. Error bars show 95% bootstrap confidence intervals over questions.
Figure 8: Aggregation fails to extract truth in verifier-absent regimes. Forecasting and commonsense benchmarks show no consistent gains from aggregation over single-sample baselines, even as agreement increases. Error bars show 95% bootstrap confidence intervals.
Benchmark
Flip Rate
Questions × Models
Predict-the-Future
3.8%
500
BoolQ
3.2%
500
HLE
2.3%
175
Com2Sense
1.5%
400
Overall
2.9%
1,575
Appendix
Table 4: Plurality answer flip rate between T=0.7 and T=1.0 . Low flip rates indicate that temperature variation does not induce the error independence required for effective aggregation.