Specialist Disagreement as a Hallucination Signal That Survives Chain-of-Thought

Asking a language model to "think step by step" (chain-of-thought prompting) makes it more accurate. It also makes its wrong answers look confident. The step-by-step text pushes probability mass onto whatever conclusion the chain reaches, so the usual warning signs of a fabricated answer fade: high output entropy, low token probability, hedged self-ratings. Chain-of-Thought Prompting Obscures Hallucination Cues in Large Language Models found that hallucination-detection AUROC fell in 465 of 768 configurations under chain-of-thought. This matters most in specialist domains such as board-level medicine, theorem-heavy STEM and appellate law. There, models lack the underlying rule or fact, and a fluent chain becomes a confident confabulation.

We propose to recover that lost signal from a different source: disagreement among reasoning chains that the same model writes while playing different specialists. The two-sentence version is this. When a model truly knows something, a cardiologist persona, a pharmacologist persona and a radiologist persona reason their way to the same claims. When the knowledge is missing, each persona fills the gap with a different plausible invention. So claim-level disagreement across specialist chains should flag hallucinations that a single chain hides. The same agreement structure can then drive the final synthesis: it keeps a well-supported minority view instead of averaging it away, and drops a fabricated claim instead of passing it along.

Core Claim

Under chain-of-thought, a hallucination detector built on claim-level cross-perspective support outperforms two alternatives in domain-knowledge-gap regimes. The first alternative is single-chain confidence signals. The second is the obvious stronger competitor: disagreement across temperature-sampled chains at matched token budget. The claim-level detector is the fraction of other specialist chains that entail, contradict or ignore each atomic claim in a chain. We further predict that its advantage concentrates on items where the model demonstrably lacks the governing principle. The mechanism is that persona conditioning sends generation through different regions of parametric knowledge. MedAgents argues that role-play draws out latent expert knowledge. If that is right, then where knowledge is absent, persona chains should make errors that are less correlated with each other than samples drawn from one neutral prompt.

This claim is falsifiable and could fail. If matched-budget self-consistency disagreement detects errors just as well, the conclusion is that diversity of any kind suffices and personas add nothing. We would report that as the result rather than argue around it.

Relation to Prior Work

Three papers motivate the gap. The chain-of-thought masking study above diagnoses the detection breakdown on open-weight models and general QA, but proposes no remedy. MedAgents runs a five-stage virtual tumor board, with expert gathering, analyses, summary, consultation and decision. It reaches 83.7% on MedQA against 82.9% for few-shot chain-of-thought with self-consistency. However, it uses agreement only to pick an answer. It never measures whether agreement predicts correctness, and never checks whether peers reinforce one another's fabrications. Chain-of-Discussion runs summarize–criticize–revise loops across several models for Chinese civil-law QA. It reports that a target model lacking domain competence absorbs peer hallucinations, and it relies on a fixed majority threshold. We reuse the multi-specialist generation of MedAgents. We replace its voting and Chain-of-Discussion's threshold with an explicit claim-by-perspective support matrix that serves as both detector and aggregation rule.

Two papers constrain the design. VITAL shows that a multi-LLM collaboration system, ModPlural, often loses to direct prompting in healthcare because its summarizer washes out minority views. Our aggregator therefore must not summarize freely. It may only keep, flag or drop claims according to their support. TheoremQA shows that pasting a theorem into the prompt barely helps, because models cannot apply rules they are handed. This rules out "just inject the rule" as the fix, and it gives us a way to label items by knowledge state.

Two papers supply methods. Do Language Models Know When They're Hallucinating References? shows that indirect, open-ended questioning exposes inconsistency in fabricated references. Our cross-examination step applies the same idea between personas, inside reasoning chains. Detecting Hallucination and Coverage Errors in Retrieval Augmented Generation for Controversial Topics defines hallucination and coverage errors for pro/con synthesis and trains judges on synthetic corruptions. We extend both error types from two-sided debates to multi-specialist technical reasoning. Finally, Prompt engineering in consistency and reliability with the evidence-based guideline for LLMs found that a single prompt simulating an expert discussion (Reflection of Thoughts) gave the highest guideline consistency, 62.9% with GPT-4. That makes it a cheap competitor we must beat.

Research Program

For each question, the model first names $K$ relevant specialists, following MedAgents' expert-gathering step. It then writes one chain-of-thought and answer per specialist. A claim extractor splits each chain into atomic claims. For every claim $c$ from perspective $i$, each other perspective $j$ is asked indirectly whether $c$ holds, without seeing $i$'s answer. The verdict is entailed, contradicted, or not addressed. The support score is

$$s(c) = \frac{1}{K-1}\sum_{j \ne i} \big(\mathbb{1}[j \text{ entails } c] - \mathbb{1}[j \text{ contradicts } c]\big),$$

which ranges from $-1$ to $1$. A chain's hallucination score is the minimum support over the claims its final answer depends on. The question-level score combines this with the spread of final answers.

Aggregation follows the matrix directly. Claims with positive support are kept. Claims that some perspectives support and none contradict are kept but marked as contested. This preserves the minority nuance that VITAL found summarizers erase. Claims that are contradicted and unsupported are dropped, which blocks the contagion seen in Chain-of-Discussion. If the support of the winning answer falls below a threshold tuned on a development split, the system abstains.

Every item is also labeled by knowledge state. The model is asked, with no chain-of-thought, to state the governing theorem, guideline or mechanism, and the statement is graded against the reference. This uses TheoremQA's theorem annotations, and on MedXpertQA a specialty-specific rubric graded by an LLM judge with human spot checks. The labels let us test whether the persona advantage sits where knowledge is missing.

Baselines

The naive reference point is direct answering without chain-of-thought, scored by its own token probability. MedAgents reports standard zero-shot prompting at 73.0% accuracy on MedQA. This reference tells us how much detection signal existed before chain-of-thought hid it.

The single-chain chain-of-thought baselines use the detectors from the masking study: perplexity, predictive entropy, internal-state probes and verbalized certainty. The masking study found these degrade under chain-of-thought, sometimes below plain perplexity. Beating them is necessary but not sufficient.

The strongest competitor for detection is self-consistency disagreement: $K$ temperature-sampled neutral chains, with answer and claim disagreement computed by exactly the same extractor and cross-check pipeline. It is the strongest competitor because it already gives multi-sample disagreement at the same cost, and on MedQA it nearly matches MedAgents in accuracy (82.9% vs 83.7%). It will receive identical token budgets, the same $K$, and its temperature tuned over the same grid we use for persona count on the development split.

The strongest existing system for accuracy is MedAgents itself, rerun on our models. The reported reference points are 83.7% on MedQA (MedAgents), 49.89% for o1 on MedXpertQA (MedXpertQA), and 51.0% for GPT-4 with Program-of-Thought on TheoremQA. We found no reported result on Lacuna for hallucination-detection AUROC on MedQA, MedXpertQA or TheoremQA under any method, so the detection comparison will establish its own numbers.

We also include Reflection of Thoughts as a single-prompt simulated discussion, and Chain-of-Discussion's fixed-threshold critique as the existing aggregation rule.

Two ablations isolate the key ingredient. The first, shuffled personas, assigns specialist labels irrelevant to the question. It tests whether relevance or mere prompt variation drives the effect. The second, answer-only support, drops claim-level cross-examination and keeps only final-answer agreement. It tests whether the claim matrix earns its cost.

Beating the strongest competitor means two things. Persona support must exceed matched self-consistency disagreement in AUROC for incorrect-answer detection, with a paired bootstrap confidence interval excluding zero, on the knowledge-gap subset of MedXpertQA and TheoremQA. And accuracy must not fall below MedAgents.

Why not just…?

Why not just use self-consistency disagreement? This is the real threat. Sampling many chains and measuring disagreement already escapes the single-chain masking effect. Nothing in the file shows that personas decorrelate errors better than temperature does. The matched-budget comparison above settles it. If the two methods tie, the paper's finding becomes "diversity, not expertise, restores the signal," and the aggregation contribution stands on its own.

Why not just answer with chain-of-thought but score confidence from a direct, no-chain answer? This decoupling sidesteps the masking effect cheaply. It cannot see claim-level fabrications inside the chain, and in knowledge-gap items the direct answer may be confidently wrong too. We include it as a baseline. If it matches our detector on multiple-choice items, persona support must justify itself on the open-ended claim-level tasks.

Why not just simulate the expert panel in one prompt? Reflection of Thoughts does exactly this, cheaply. In a single context, however, later "experts" condition on earlier ones, which is the compounding failure that Holt et al. describe and the contagion Chain-of-Discussion reports. Running independent chains is the property we test. The single-prompt variant is run at equal tokens to measure what independence buys.

Why not just retrieve the rule or call a tool? For lookup facts, tools win outright. Deficiency of Large Language Models in Finance reports 100% on stock-price queries with tools. But TheoremQA shows that supplying the theorem does not teach the model to apply it, and Chain-of-Discussion shows that retrieval returns plausible distractors. Retrieval also offers no detection signal of its own. We run a theorem-augmented arm on TheoremQA to check whether the knowledge-gap advantage persists once the rule is in context.

Proposed Experiments

The main study uses open-weight models whose token probabilities are visible, matching the masking study: Llama-3.1-8B, Llama-3.1-70B and DeepSeek-R1-Distill-Llama-8B. It covers MedQA, the MedXpertQA text subset (2,455 questions across 17 specialties) and TheoremQA (800 questions over 350 theorems). We also replicate on MMLU and TruthfulQA, to confirm we reproduce the masking effect before claiming to fix it. Metrics are AUROC and expected calibration error for detecting incorrect answers, selective accuracy at fixed coverage, and plain accuracy, each reported separately by knowledge-state stratum.

For the aggregation claim, we use the 80-pair ProCon set from the coverage-error paper and a sample of VITAL Overton items. On these we measure hallucination and coverage errors of the synthesized answer against free summarization (ModPlural-style) and Chain-of-Discussion thresholds, with human annotation on a subset. A secondary probe on CEBQA tests whether contested-claim marking reduces adoption of distractor statutes.

Three checks decide the outcome. First, persona support versus matched self-consistency on the knowledge-gap stratum is the central result. Second, the shuffled-persona ablation should lose most of the advantage if expertise rather than noise is responsible. Third, the advantage should shrink on items where the model can state the governing principle. If it does not, the knowledge-gap explanation is wrong even if the detector works.