Language models are now routinely graded by other language models. A strong model is asked to read another model's step-by-step reasoning and say whether each step is right. This "LLM-as-a-judge" setup is how many agent pipelines, retrieval systems, and reasoning benchmarks are audited when there are no gold intermediate answers. There is a specific failure this setup may hide. Suppose a question is answered over several dependent steps (a multi-hop question), and the provided documents state a fact that runs against a social stereotype, for example that a 71-year-old is a company's lead mobile developer. Generators are known to quietly overwrite such facts with the stereotypical version, and to do so confidently. If the judge carries the same stereotype, it may read the overwritten step as plausible and approve it. The audit would then fail exactly where the generator fails.
We propose to measure this "shared-prior blind spot" directly. We build multi-hop questions about fictional people in which one hop's fact is stereotype-consistent, stereotype-inconsistent, or neutral. We plant matched errors into reasoning chains: one error replaces the stated fact with the stereotypical value, the other replaces it with an equally wrong neutral value. We then ask whether step-level LLM judges catch the two kinds of error at the same rate, whether the judges locate them at the right step, and whether the gap grows with the number of reasoning steps that follow the error. Because the data are constructed, every step has a gold answer. The judge can therefore be scored against ground truth, not against another model.
Core Claim
The claim is narrow and testable. When a reasoning chain contains an error that swaps a provided, stereotype-inconsistent fact for its stereotypical counterpart, step-level LLM judges detect and localize it less often than a position-matched error of the same factual severity that carries no stereotype. The gap widens when the error sits early in the chain, because later steps reason consistently from the wrong premise and make the chain look coherent.
We define the blind-spot gap at hop position $k$ as
$$G_k = R^{\text{neutral}}_k - R^{\text{stereo}}_k,$$
where $R^{\text{neutral}}_k$ and $R^{\text{stereo}}_k$ are the judge's step-level detection recall for planted neutral and stereotype-congruent errors at hop $k$. The two error types are matched on hop, chain length, and factual wrongness relative to the context. A positive $G_k$ that is significant across demographic axes, and that does not shrink when the judge is shown the source documents, is the finding. If $G_k$ is near zero everywhere, that is also useful: step-level judges would then be safe for this kind of auditing.
Why This Is Not Already Done
Four lines of work come close, and each stops short in a different place.
The motivating result is Exploring Causal Effect of Social Bias on Faithfulness Hallucinations in Large Language Models. It uses pro-, anti-, and non-stereotype interventions on its Bias Intervention Dataset (BID) and finds significant causal effects on context-faithfulness errors in 85 of 105 settings. It names these errors "unfairness hallucinations" and shows they come with higher model confidence than neutral errors. That work studies single-step generation only. It never asks whether a downstream judge catches these errors, and it never places them inside dependent reasoning chains.
CofCA: A STEP-WISE Counterfactual Multi-hop QA benchmark supplies our construction method. It replaces real entities with fictional ones to separate reasoning from memorization, and reports that GPT-4 produces fully correct chains on only 36.3% of items. Its counterfactuals are knowledge edits, not social-identity edits, and it grades with exact match, not with judges. ♫ MuSiQue: Multihop Questions via Single-hop Question Composition supplies the bottom-up composition and the filtering that stops models from skipping hops.
On the judge side, The Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge swaps provenance and date labels on fixed text. It finds verdict shifts of up to +30% with a 0% rate of judges acknowledging the cue. BiasScope: Towards Automated Detection of Bias in LLM-as-a-Judge Evaluation injects biases automatically and drives GPT-4o to a 74.7% error rate on JudgeBench-Pro. Notably, when BiasScope was seeded with social biases, its search drifted toward cognitive biases. Neither paper perturbs social identity inside reasoning, and both judge flat final answers.
The step-level judging methods we test against are AgentEval: DAG-Structured Step-Level Evaluation for Agentic Workflows with Error Propagation Tracking, which grades workflows as dependency graphs with GPT-4o, and AgentHallu: Benchmarking Automated Hallucination Attribution of LLM-based Agents, which benchmarks error localization. Both treat the judge as a neutral referee, and neither contains demographic content. The closest single paper is therefore the BID study, and the difference is that it measures the generator while we measure the auditor, inside multi-step chains, against gold labels.
Research Program
Data construction. We compose two-, three-, and four-hop questions MuSiQue-style from single-hop facts about fictional people and organizations, as in CofCA, so that no answer can come from memory. Exactly one hop, the "identity hop," states an attribute tied to a person whose identity is marked along one of BID's axes: age, gender, disability, religion, or socioeconomic status. Each base item gets three versions in which that fact is pro-stereotype, anti-stereotype, or non-stereotype, following BID's design. Following Counter-GAP: Counterfactual Bias Evaluation through Gendered Ambiguous Pronouns, each item also gets within-group identity swaps, for example a different name of the same coded group. These let us subtract the noise that any name change induces from the effect of changing groups. We vary where the identity hop sits (first, middle, last), so we can test depth. Anchor templates will be reviewed by annotators to confirm that the stereotype direction is the intended one.
Generator study. We run open and proprietary models with plain chain-of-thought and with interleaved retrieval. The interleaved setting uses IRCoT, from Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions, where each reasoning sentence becomes the next search query. Retrieval matters because a stereotype-overwritten fact at hop one can change what is fetched at hop two. Gold intermediate answers let us label each hop as correct, as an unfairness hallucination (the stereotypical value replaces the stated one), or as another error. They also let us estimate how often an error at hop $k$ propagates to the final answer. This half replicates BID's effect in chains and produces naturally occurring errors.
Judge study. This is the main contribution. We take correct chains and plant matched errors at controlled hops. The stereotype-congruent version substitutes the stereotypical value. The neutral version substitutes a wrong value of the same type that carries no stereotype. Downstream steps are rewritten so that they reason consistently from the planted premise, as a real cascading error would. Judges grade every step under three conditions: without the source documents, with the source documents, and in AgentEval's dependency-aware format. We record detection, step localization, the assigned error category, and whether the written rationale mentions identity.
A second judge measurement holds the chain text fixed and swaps only the person's group. We then compute a noise-corrected verdict shift
$$\Delta V = V_{\text{across}} - V_{\text{within}},$$
where $V_{\text{across}}$ and $V_{\text{within}}$ are the rates at which the judge's step verdict flips under across-group and within-group identity swaps. This adapts Counter-GAP's inconsistency difference from coreference to judging. Finally, we validate the planted-error findings on the generator's naturally occurring unfairness hallucinations, so the conclusions do not rest only on synthetic edits.
Baselines
The naive reference point is end-to-end counterfactual fairness: check only whether the final answer matches gold across identity versions. AgentEval reports that end-to-end checking reaches failure detection recall of 0.41, against 0.89 for step-level evaluation. It is included because it is what most fairness audits do. It will be run on exactly the same items, and the comparison is how many planted stereotype errors it misses because the chain still lands on the right answer.
The strongest existing judge is AgentEval's dependency-aware step-level GPT-4o judge. Its paper reports Cohen's κ = 0.84 with human experts and 72% root-cause accuracy against an 81% human ceiling. We reimplement its step rubric and propagation tracking on our chains, using the same judge models and the same prompt-tuning budget as every other judge condition, with tuning done on a held-out split that contains only neutral errors. For localization we also compare with AgentHallu-style step-by-step attribution prompting. AgentHallu reports 41.1% step localization accuracy for Gemini-2.5-Pro, falling to 11.4% on trajectories of 11 or more steps. "Beating" these baselines is not the goal here. The result that matters is whether they show a significant positive $G_k$ that their headline accuracy hides.
A third baseline is uncertainty-based detection: flag steps whose token-level confidence is low. BID reports that unfairness hallucinations are more confident than neutral errors but gives no detection score for this setting, so no reported result was found on Lacuna for uncertainty-based detection of unfairness hallucinations in chains. We will measure it on open-weight generators.
Two baselines target the judge's bias directly. The first is statistical correction in the style of HiBayES, from Quantifying Biases in LLM-as-a-Judge Evaluations, which models judge scores with Bayesian regression and subtracts the estimated shifts. On MT-Bench it raised simulated agreement from α = −0.2 to α = 0.7. We fit it with demographic group and stereotype condition as predictors. The second is the key ablation: a judge given the gold source sentence for each hop. If grounding removes $G_k$, the blind spot comes from missing evidence. If it persists, the judge's prior is overriding evidence it can see.
Why not just…?
Why not just run BID-style single-step probes on each hop and multiply the error rates? This is the most serious threat to the generator half of the study. If identity effects act independently at each hop, chains add nothing new. We test it directly by predicting chain-level unfairness-hallucination and propagation rates from single-hop rates measured on the same facts, then comparing those predictions with observed chain rates. If the prediction holds within confidence intervals, we report that and shift the weight of the paper onto the judge half, which single-step probes cannot address at all.
Why not just score intermediate hops by string match against gold, with no LLM judge? On constructed data we do exactly this, because it is how we score the judges. The question we study is what happens in deployment, where gold intermediate answers do not exist and judges are the auditing tool. AgentEval and AgentHallu describe that setting. String matching cannot be the deployed method, so the useful question is how far the judge falls from it on stereotype errors.
Why not just read the judge's rationale for signs of demographic reasoning? The Silent Judge found a 0% cue acknowledgment rate even when verdicts shifted by up to 30%. Biases in the Blind Spot: Detecting What LLMs Fail to Mention reports biases that never appear in chain-of-thought. We still record acknowledgment so the claim is tested in our setting, but we expect rationales to be uninformative.
Why not just apply the Silent Judge's swap to final answers? That probe measures whether a verdict shifts. It cannot measure missed errors, which require a planted ground-truth error, and it cannot measure dependence on depth. If flat judging shows the same gap as step-level judging, that is reported as a finding, not as a failure of the design.
Decisive Checks
The central test is whether $G_k > 0$ for at least two of the five demographic axes under the AgentEval-style judge. We correct for multiple comparisons across axes, hops, and judges, and use McNemar tests on matched error pairs, as BID does. The second check is the slope of $G_k$ with hop position: the depth hypothesis predicts a larger gap for first-hop errors when three or more consistent steps follow them. The third check is the grounding ablation, which separates missing evidence from prior override. The fourth is whether $\Delta V$ on fixed chains is significant after the within-group correction; without that correction, name-frequency noise alone could produce an apparent effect. The fifth is whether HiBayES-style correction closes the gap. Closing a detection gap requires changing which errors the judge finds, not shifting its scores, so we expect correction to fail. If it succeeds, that is a cheap fix worth reporting.
The study fails cleanly in two ways. If matched stereotype and neutral errors are detected at equal rates, step-level judges are safe for this use. If chain behavior matches the single-hop predictions, the multi-hop framing adds nothing for generators. Either outcome would be informative to anyone relying on LLM judges to audit fairness in reasoning systems.
References
- Exploring Causal Effect of Social Bias on Faithfulness Hallucinations in Large Language Models
- CofCA: A STEP-WISE Counterfactual Multi-hop QA benchmark
- ♫ MuSiQue: Multihop Questions via Single-hop Question Composition
- The Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge
- BiasScope: Towards Automated Detection of Bias in LLM-as-a-Judge Evaluation
- AgentEval: DAG-Structured Step-Level Evaluation for Agentic Workflows with Error Propagation Tracking
- AgentHallu: Benchmarking Automated Hallucination Attribution of LLM-based Agents
- Counter-GAP: Counterfactual Bias Evaluation through Gendered Ambiguous Pronouns
- Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
- Quantifying Biases in LLM-as-a-Judge Evaluations
- Biases in the Blind Spot: Detecting What LLMs Fail to Mention