Black-box hallucination detectors rest on one bet: when several language models are asked the same thing, true content is reproduced and fabricated content is not. A claim that many models support is trusted, and a claim few support is flagged. Social choice theory has a well-known caveat about this bet. Majority voting recovers the truth only when voters err independently. A bloc of voters who learned the same mistakes can outvote everyone else. Language models are such a bloc. They are trained on overlapping web text, and they often distill from one another. TruthfulQA showed that larger models reproduce popular human misconceptions more faithfully than smaller ones. On exactly the questions where hallucination matters most, then, model agreement may measure shared upbringing rather than truth.
This note proposes to estimate how dependent the judging models are on each other from their agreement patterns alone, with no labels, and then to aggregate their rankings so that each block of mutually dependent models gets roughly one vote. The test is built on a simple fact. If models erred independently given the truth, the matrix of pairwise agreement between models would have rank one, after a suitable centering. Structure beyond rank one is a measurable signature of shared errors. Clustering on that residual structure, then running a two-level rank aggregation in the style already used for NLP leaderboards, gives a consensus that should resist shared misconceptions. Where every model shares the misconception, it should at least flag "suspicious unanimity."
Core Claim
The claim has two parts. The first is a diagnostic. On benchmarks built around shared misconceptions, the centered agreement matrix among a pool of models will depart sharply from rank one. On benchmarks where hallucinations are idiosyncratic fabrications, it will be close to rank one. The second is a method. Aggregating per-model rankings of candidate answers in two stages, first within dependence clusters and then across them, will raise detection accuracy over flat consensus in the first regime and leave it unchanged in the second. If the diagnostic does not separate the regimes, the method has no reason to work, and that is a cheap, early way to falsify the idea.
Where This Sits
Several papers motivate the proposal by finding the gap. SelfCheckGPT samples several completions from one model and treats inconsistency with the main output as a hallucination signal, averaging similarity scores across samples. CrossCheckGPT extends this across models. It checks a target model's output against several "evidence models" and weights each evidence model by its own SelfCheckGPT score. That weighting rewards models that are individually consistent. It does nothing about two consistent models that are consistent in the same wrong way, and the file notes this as an open limitation of that work. TruthfulQA supplies the reason to expect correlated errors. On its 817 adversarial questions, the 175B model was less truthful than its 350M counterpart. That result also supplies a hopeful twist: small models can form a dissenting bloc that dependence-aware aggregation would stop drowning out.
Other work supplies the method. What are the best Systems? treats benchmark tasks as voters and systems as candidates. It replaces score averaging with Borda-approximated Kemeny consensus, where Kemeny consensus is the ranking that minimizes total pairwise disagreements with the inputs and Borda is the cheap positional approximation. It also introduces a two-level procedure that aggregates instances into tasks and then tasks into a leaderboard. We reuse that two-level structure, but our first level groups dependent judges rather than instances, and the grouping is learned from data. That paper explicitly treats voters as equally weighted and independent, which is the assumption we drop. Sync-Rank motivates the spectral view. It recovers rankings from noisy comparisons through the top eigenvector of a comparison matrix and can detect planted partial rankings. Our diagnostic follows the same spirit: the first eigenvector captures shared competence, and the next ones expose planted blocs.
Two results constrain the design. HaluEval reports that LLM judges discriminate worse when shown a correct and a hallucinated candidate side by side than when shown one alone. Each judge therefore scores every candidate in isolation, and rankings are formed afterwards from those scores, never by asking for a comparison. The survey A Survey on Rank Aggregation and Rank aggregation: New bounds for MCx document that Markov-chain aggregation rules degrade on noisy, inconsistent voters, so we restrict ourselves to Borda and approximate Kemeny.
Research Program
The setting has a pool of $m$ judge models and a set of items. Each item has several candidate answers or claims: the true and false reference answers of a TruthfulQA question, the correct and hallucinated responses of a HaluEval item, or the sentences of a WikiBio passage. Each judge scores each candidate alone for factual support, and the scores give that judge's ranking $\sigma_{ji}$ of the candidates for item $i$. Because only orderings are used, judges whose scores live on incompatible scales can be combined without calibration. That is the scale-invariance argument of Colombo et al.
The dependence diagnostic rests on one identity. Suppose each judge picks the correct side of a binary contrast with accuracy $p_j$, independently of the others given the truth. Then the centered agreement between judges $j$ and $k$ factorizes:
$$\mathbb{E}[\,2\,\mathbb{1}(j \text{ and } k \text{ agree}) - 1\,] = (2p_j - 1)(2p_k - 1).$$
Here the indicator equals one when the two judges order a candidate pair the same way, and the expectation runs over pairs. The identity follows by summing the two ways the judges can agree, both right or both wrong. For rankings we use the analogous quantity: the Kendall tau correlation between $\sigma_{ji}$ and $\sigma_{ki}$, averaged over items, which forms a matrix $A$. Under conditional independence, $A$ is approximately rank one off the diagonal. We fit the best rank-one approximation, take the residual $R = A - \hat a \hat a^\top$, and cluster judges using the leading eigenvectors of $R$. The fraction of off-diagonal energy left in $R$ becomes a scalar "dependence index" for a model pool on a benchmark.
Aggregation then runs in two levels. Within each cluster, a Borda consensus produces one ranking per cluster. Across clusters, a second Borda or approximate Kemeny step produces the final ranking, so a bloc of eight distilled siblings counts about as much as one independent small model. The method also outputs a per-item flag. An item is flagged when the flat consensus and the dependence-aware consensus disagree on the top candidate, or when the winning margin comes almost entirely from a single cluster. That flag targets the case where reweighting cannot help because every bloc shares the error: abstention is then the honest response.
The pool should deliberately mix model families, sizes from sub-billion to the largest open models available, and known distilled derivatives, so that dependence has a chance to show up. The diagnostic can then be checked against provenance metadata that the method never sees.
Baselines
The naive reference point is flat majority consensus. Every judge scores candidates in isolation, and a single Borda count over all judges picks the winner. This is the method's strongest internal competitor. It differs only in ignoring dependence, so beating it on TruthfulQA while tying it on WikiBio is the central result. Alongside it we run arithmetic averaging of raw judge scores, to confirm the rank-versus-score advantage that Colombo et al. report. In their analysis of over 270,000 scores, the system ranked third by mean aggregation on XTREME dropped to sixth under rank aggregation.
The strongest existing method is CrossCheckGPT's explicit and implicit cross-checking, with judges weighted by their SelfCheckGPT confidence, restricted to text. It uses the same judge pool, the same number of sampled passages per judge, and the same isolated-scoring prompts. CrossCheckGPT reports 98% Spearman rank correlation with human judgments on MHaluBench and 89% on AVHalluBench, but these are model-level rankings on multimodal data. The file gives no reported CrossCheckGPT number on WikiBio or TruthfulQA, so we rerun it ourselves. For sentence-level detection on WikiBio, the strongest published single-model number is SelfCheckGPT's prompt variant at AUC-PR 93.42 for non-factual sentences (SelfCheckGPT). On TruthfulQA, the reference points are the original paper's 58% truthfulness for the best GPT-3 configuration and roughly 94% for humans. No reported result was found on Lacuna for any cross-model consensus detector on TruthfulQA. For HaluEval QA, ChatGPT as a single judge reached 62.59% accuracy.
Three ablations isolate the key ingredient. One replaces learned residual clusters with clusters defined by model family. A second uses random clusters of matched sizes. A third keeps the clusters but replaces the second-level rank aggregation with score averaging. Every variant gets the same queries per judge and the same total compute, and no variant sees labels. Beating the baselines means a statistically reliable gain over flat Borda and CrossCheckGPT in top-candidate accuracy on TruthfulQA, measured by a paired bootstrap over questions, together with no loss on WikiBio AUC-PR.
Why not just…?
Why not just pick judges from different model families? This is the most serious threat, because provenance is a cheap proxy for dependence. Shared misconceptions, though, come from shared human text, which crosses family boundaries. Distillation also links models whose provenance labels look unrelated. The family-cluster ablation settles the question directly. If it matches learned clusters, the honest conclusion is that the diagnostic is a validation tool rather than a better aggregator, and we will report it that way.
Why not just use a few labels to learn judge weights? A RankBoost-style supervised combiner (An Efficient Boosting Algorithm for Combining Preferences) trained on a small labeled set could learn to down-weight the misconception bloc. We will run it at label budgets of tens to hundreds of items. Our method's claim is limited to the label-free regime. If a combiner with fifty labels dominates and transfers across TruthfulQA categories, that limits the contribution, and we would say so.
Why not just ask one strong model? HaluEval shows single-judge detection near chance, at 58.53% on summarization and 62.59% on QA for ChatGPT. TruthfulQA's inverse scaling suggests that the biggest single judge may be the one most committed to popular errors. The single strongest pool member, run as a judge with the same prompts, is included as a baseline.
Why not just use token probabilities? SelfCheckGPT found that sampled black-box consistency outperformed grey-box probability signals on WikiBio, and API models often hide these probabilities. Probability-based ranking is also exactly the signal a memorized misconception makes confident.
Proposed Experiments
The first and decisive check is the diagnostic itself. We compute the dependence index for the same judge pool on TruthfulQA, HaluEval QA, and WikiBio GPT-3 (238 biographies, 1,908 sentences). The prediction is a large residual on TruthfulQA and a near-rank-one matrix on WikiBio. If the residual is no larger on TruthfulQA, the premise fails and the method is abandoned. A synthetic control precedes this. We plant a bloc of judges that copy a shared corrupted ranking on a known fraction of items and verify that residual clustering recovers the bloc, measured against the planted membership as noise grows.
The main experiment compares dependence-aware two-level aggregation against all baselines above. The metrics are top-candidate accuracy on TruthfulQA, with judged truthfulness checked using the TruthfulQA authors' automated evaluator, which reported 90–96% agreement with humans; accuracy on HaluEval QA and dialogue; and AUC-PR plus passage-level Pearson correlation on WikiBio. A secondary analysis tests the TruthfulQA-derived hypothesis that small models form a dissenting cluster. We check whether the gain concentrates on questions where the small-model cluster disagrees with the large-model cluster, and whether the suspicious-unanimity flag gives better selective accuracy, measured as accuracy against coverage, than flagging by low consensus margin alone.
The last check addresses robustness. We add near-duplicate judges, meaning fine-tunes of a single base model, to the pool and measure how flat consensus and the proposed method change. Flat consensus should drift toward the duplicated model's errors in proportion to the number of copies. The dependence-aware consensus should barely move. That invariance to cloning a voter is the property the whole proposal is built to deliver.
References
- SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
- CrossCheckGPT: Universal Hallucination Ranking for Multimodal Foundation Models
- TruthfulQA: Measuring How Models Mimic Human Falsehoods
- What are the best Systems? New Perspectives on NLP Benchmarking
- Sync-Rank: Robust Ranking, Constrained Ranking and Rank Aggregation via Eigenvector and SDP Synchronization
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- A Survey on Rank Aggregation
- Rank aggregation: New bounds for MCx
- An Efficient Boosting Algorithm for Combining Preferences