Reliability of LLM Judges for Evaluating Entity Alignment¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/vaibhavalakshmiravideshik/llm-as-a-judge-entity-alignment
Area: LLM Reasoning
Keywords: LLM Judge, Entity Alignment, Anchor Bias, Counterfactual Evaluation, Frontier Model Paradox
TL;DR¶
This paper exposes the severe vulnerability of LLM judges to "anchor bias" in entity alignment evaluation, where displaying system decision labels causes correctness discrimination to catastrophically invert (J-ROC-AUC dropping to 0.12–0.87) with stronger frontier models rationalizing labels even more aggressively (the frontier model paradox); withholding system labels under a label-free prediction protocol completely restores near-ceiling discrimination capability (0.93–1.00).
Background & Motivation¶
Entity Alignment (EA) seeks to determine whether entities from disparate, heterogeneous knowledge graphs correspond to the identical real-world concept, forming the foundational pillar for knowledge graph integration, ontology merging, and cross-domain data federation. While modern EA architectures have advanced from translation-based vector embeddings to complex neuro-symbolic and LLM-guided systems, a persistent structural bottleneck hampers progress: verifying candidate alignments across millions of relational triples demands exhaustive domain-expert annotation, rendering large-scale automated benchmarking practically unaffordable. Evaluating EA systems via reference-free LLM-as-a-judge paradigms presents an enticing path forward, yet whether frontier language models possess the disciplined ability to adjudicate structured factual evidence independently without leaning on pre-trained surface correlations has remained unverified.
The fundamental breakdown of this evaluation setup arises when the candidate prediction label generated by an EA system (MATCH or NON-MATCH) is placed directly in the prompt context of the LLM judge. Under this formulation, the task presented to the judge imperceptibly mutates from rigorous evidence auditing into post-hoc rationalization. Consider a representative pair from DBpedia and Wikidata: both entities share the identical surface string "The Fast and the Furious" and the same film director, yet Entity A denotes the 2001 standalone feature film while Entity B denotes the entire media franchise. When an EA system accurately predicts NON-MATCH, an LLM judge evaluating this decision assigns a dismal justification score of 3/10 because it anchors on the shared lexical surface tokens rather than discerning the underlying ontological distinction. The visible system label functions as a cognitive anchor, inducing confirmation bias and leading the judge to systematically validate incorrect decisions while penalizing correct ones, creating a circular auditing loop.
The objective of this work is to establish the first rigorous causal auditing benchmark for LLM judges in entity alignment, formally categorizing and quantifying anchor bias across three frontier models (GPT-4o-mini, GPT-4o, and Claude Opus 4.7), four EA systems and ablations, and three benchmark datasets. Through a combination of systematic input perturbation tests, multi-prompt meta-evaluation, and counterfactual label-flip interventions, the authors demonstrate that conventional chain-of-thought engineering is utterly insufficient to neutralize this structural vulnerability. Core idea: LLM judge failures in entity alignment do not stem from reasoning inability but from structural anchor bias induced by visible decision labels—a failure mode amplified rather than mitigated by model capability (the frontier model paradox); physically eliminating system labels via a label-free prediction protocol is the sole effective mechanism to unlock genuine entity discrimination.
Method¶
Overall Architecture¶
The evaluation framework comprises an end-to-end multi-stage diagnostic pipeline spanning perturbation diagnostics, multi-prompt meta-evaluation, counterfactual label-flip interventions, and evaluation across a newly curated heterogeneous biomedical benchmark. For any candidate entity pair \((e_1, e_2)\), its surrounding graph context, and the candidate system decision \(d \in \{\text{MATCH}, \text{NON-MATCH}\}\), the pipeline first isolates lexical name reliance via controlled perturbations. It then evaluates discrimination quality across three distinct prompt interaction paradigms, applies counterfactual interventions holding evidence constant while flipping the decision label to isolate direct causality, and tests generalizability across a multi-schema biomedical ontology.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Graph Entity Pair & System Decision<br/>(e1, e2, Context, Decision d)"] --> B["Perturbation Bias Diagnostics<br/>PBI / JSR ρ / MRI Measure Surface Name Shortcuts"]
B --> C["Three-Prompt Meta-Evaluation Paradigm<br/>Prompts I-III Expose Discrimination Collapse"]
C --> D["Counterfactual Label-Flip Interventions<br/>MAS / FSR / LFR Reveal Frontier Model Paradox"]
D --> E["Heterogeneous Biomedical Benchmark<br/>MeSH × SNOMED CT 15K Cross-Schema Stress Test"]
E --> F["Auditing Conclusion & Protocol Redesign<br/>Abolish Label Display / Adopt Label-Free Prediction"]
Key Designs¶
1. Perturbation Bias Diagnostics: Disentangling Lexical Shortcuts from Structural Reasoning To diagnose whether an LLM judge evaluates graph topological evidence or merely defaults to memorized pre-training associations of entity names, the authors established a suite of behavioral perturbation strategies, prominently contrasting identity anonymization (P1), 30% and 50% evidence suppression (P2, P3), identifier removal (P6), and order randomization (P5, serving as a negative control). To quantify the sensitivity and consistency of judge degradation, three complementary diagnostic metrics were formalized: the Perturbation Bias Index (PBI) measuring the average raw score drop: $\(\text{PBI} = \bar{s}_{\text{original}} - \bar{s}_{\text{perturbed}}\)$ the Judge Score Rank Correlation (JSR \(\rho\)) computing the Spearman rank correlation of scores before and after perturbation to evaluate macro-level ranking stability; and the Massive Rank Inversion rate (MRI) measuring the percentage of pairs undergoing catastrophic score collapse exceeding 2 points: $\(\text{MRI} = \frac{|\{i : s_i^{\text{original}} - s_i^{\text{perturbed}} > 2\}|}{n} \times 100\%\)$ This diagnostic suite reveals that whenever entity names are obscured, judges lacking dense pre-training knowledge experience dramatic ranking scrambling, demonstrating heavy reliance on lexical recognition shortcuts.
2. Three-Prompt Meta-Evaluation Paradigm: Exposing Discrimination Inversion Under Label Exposure To dissect the structural vulnerability of label exposure, the authors designed three prompt conditions: Prompt I (Evidence-Locked Evaluation) enforces an exhaustive 8-step chain-of-thought protocol explicitly instructing the judge to discard pre-trained memory and justify the visible decision exclusively from provided evidence strings; Prompt II (Minimal Evaluation) presents the raw entity evidence alongside the system decision without guidance, mirroring default practitioner behavior; and Prompt III (Label-Free Alignment Prediction) withholds the system decision entirely, prompting the LLM to independently output its alignment confidence score from 0 to 10. System correctness discrimination was formalized via J-ROC-AUC, computing \(P(s_{\text{correct}} > s_{\text{incorrect}})\) under Prompts I–II, alongside the empirical Score Gap: $\(\Delta = \bar{s}_{\text{MATCH}} - \bar{s}_{\text{NON-MATCH}}\)$ The empirical results demonstrate that across Prompts I and II, J-ROC-AUC frequently collapses below the 0.5 chance threshold down to 0.12–0.53, accompanied by negative score gaps where judges systematically assign higher justification scores to incorrect decisions. Conversely, withholding labels under Prompt III immediately elevates J-ROC-AUC to 0.93–1.00, proving that complex prompt instructions fail to prevent post-hoc label validation.
3. Counterfactual Label-Flip Interventions and the Frontier Model Paradox Because correlation between visible labels and assigned scores does not formally establish causality, the authors introduced counterfactual interventions: holding all structural graph evidence and entity descriptions strictly identical, the visible system label was artificially flipped (MATCH \(\leftrightarrow\) NON-MATCH). The causal impact was captured using three metrics: Mean Absolute Shift (MAS) assessing average score displacement \(\frac{1}{n}\sum_{i=1}^n |s^{(i)} - \bar{s}^{(i)}|\); Flip Sensitivity Rate (FSR) measuring the percentage of instances with score shifts exceeding 2 points: $\(\text{FSR} = \frac{|\{i : |s^{(i)} - \bar{s}^{(i)}| > 2\}|}{n} \times 100\%\)$ and Label Following Rate (LFR) calculating the fraction of shifts moving directionally toward the newly flipped label. Across all evaluations, FSR exceeded 50% (ranging from 53% to 99%), definitively establishing visible labels as the primary causal driver of scoring. Crucially, this uncovered the "Frontier Model Paradox": the most capable model (Claude Opus 4.7) demonstrated the highest label sensitivity (FSR 83%–99%, MAS up to 8.21), whereas the smaller GPT-4o-mini showed significantly higher resistance (FSR 53%–64%). Stronger reasoning capabilities paradoxically amplify post-hoc rationalization rather than bolstering independent judgment.
4. Heterogeneous Biomedical Benchmark and Domain-Asymmetric Failure Modes To overcome the artificial simplicity of existing homogeneous Wikipedia-derived benchmarks (such as DBpedia–Wikidata/YAGO), the authors created and publicly released BIO (MeSH–SNOMED CT 15K), a real-world biomedical alignment benchmark. MeSH is organized around hierarchical medical subject headings and alphanumeric tree numbers, whereas SNOMED CT is structured as a description-logic clinical ontology with dense concept relationships and many-to-many granularities. This ontological heterogeneity uncovered an asymmetric failure mechanism: while general-domain judges exhibited score variance compression under minimal prompting (PDV ratios 0.21–0.81) due to strong pre-trained name familiarity, biomedical judges exhibited severe score variance inflation (PDV ratios 1.17–1.35) and catastrophic identity anonymization collapse (MRI 61%–69%), demonstrating that absence of pre-trained entity familiarity destabilizes LLM judge scoring completely.
Benchmark & Evaluation Protocol¶
The empirical evaluation was conducted across three datasets: DW (DBpedia × Wikidata, distinctive entity names), DY (DBpedia × YAGO, near-identical entity names), and BIO (MeSH × SNOMED CT, cross-schema biomedical benchmark). Target alignment systems spanned EasyEA (LLM-augmented candidate selection) and its ablated variant EasyEA-EmbOnly, alongside NeuSymEA (neuro-symbolic variational inference) and NeuSymEA-NoRefine. Three closed-source frontier judges were assessed: GPT-4o-mini, GPT-4o, and Claude Opus 4.7 (evaluated at temperature = 0 in single continuous sessions). To externally validate automated findings, a blinded two-annotator human evaluation was executed on 102 pairs across four negative-mining strategies (Cohen's \(\kappa = 0.902\), gold-label consensus accuracy 91.3%), confirming that human-LLM scoring disagreement is intensely concentrated on label-exposed false validations.
Key Experimental Results¶
Main Results: Meta-Evaluation Discrimination Across Prompts (EasyEA & NeuSymEA)¶
The table below presents correctness discrimination performance (J-ROC-AUC) and Score Gap (\(\Delta = \bar{s}_M - \bar{s}_{NM}\)) across all three prompting conditions (\(n=500\) pairs per condition). Prompts I and II reveal pervasive discrimination collapse and inversion, whereas Prompt III unlocks near-ceiling alignment capability:
| Dataset | EA System | Judge Model | Prompt I (CoT) J-ROC-AUC | Prompt I Score Gap | Prompt II (Minimal) J-ROC-AUC | Prompt II Score Gap | Prompt III (Predictor) J-ROC-AUC | Prompt III Score Gap |
|---|---|---|---|---|---|---|---|---|
| DW | EasyEA | GPT-4o-mini | 0.528 | +0.14 | 0.244 | -0.79*** | 0.974 (F1-best) | +8.15*** |
| DW | EasyEA | GPT-4o | 0.445 | -0.32 | 0.235 | -0.92*** | 0.969 (F1-best) | +8.09*** |
| DW | EasyEA | Claude Opus 4.7 | 0.342 | -0.76*** | 0.494 | -0.10 | 0.972 (F1-best) | +7.80*** |
| DW | NeuSymEA | GPT-4o-mini | 0.355 | -0.90*** | 0.145 | -1.07*** | 0.981 | +8.25*** |
| DW | NeuSymEA | GPT-4o | 0.251 | -1.21*** | 0.165 | -1.00*** | 0.995 | +8.30*** |
| DW | NeuSymEA | Claude Opus 4.7 | 0.189 | -1.25*** | 0.451 | -0.12*** | 0.995 | +8.32*** |
| DY | EasyEA | GPT-4o-mini | 0.567 | +1.80** | 0.712 | +3.98*** | 0.527 | +2.25 |
| DY | EasyEA | GPT-4o | 0.783 | +4.05*** | 0.750 | +4.79*** | 0.468 | +2.01 |
| DY | EasyEA | Claude Opus 4.7 | 0.766 | +4.37*** | 0.866 | +6.77*** | 0.507 | +1.69 |
| DY | NeuSymEA | GPT-4o-mini | 0.509 | +0.03 | 0.197 | -0.75*** | 1.000 | +9.13*** |
| DY | NeuSymEA | GPT-4o | 0.243 | -1.18*** | 0.123 | -1.09*** | 0.999 | +8.45*** |
| DY | NeuSymEA | Claude Opus 4.7 | 0.129 | -1.60*** | 0.477 | -0.02 | 1.000 | +7.93*** |
| BIO | EasyEA | GPT-4o-mini | 0.477 | -0.11 | 0.341 | -0.75*** | 0.931 | +6.12*** |
| BIO | EasyEA | GPT-4o | 0.423 | -0.42** | 0.295 | -0.70*** | 0.945 | +6.20*** |
| BIO | EasyEA | Claude Opus 4.7 | 0.223 | -1.02*** | 0.457 | +0.25 | 0.948 | +6.00*** |
(Note: *** \(p < 0.001\), ** \(p < 0.01\); on DY EasyEA, near-identical entity names leave zero discriminative signal without structural graph refinement, limiting Prompt III recovery)
Ablation & Causal Analysis: Counterfactual Label-Flip Intervention (Prompt I)¶
The table below reports causal sensitivity metrics under Prompt I with fixed evidence and counterfactually flipped system decision labels (\(n=100\), 50 MATCH + 50 NON-MATCH). The data illustrates the monotonic rise of label sensitivity with model scale:
| Dataset | Judge Model | EA System | Mean Absolute Shift (MAS) | Flip Sensitivity Rate (FSR%) | 95% Bootstrap CI | Label Following Rate (LFR%) |
|---|---|---|---|---|---|---|
| DW | GPT-4o-mini | EasyEA | 3.55 | 64% | [55%, 73%] | 17% |
| DW | GPT-4o | EasyEA | 4.66 | 71% | [62%, 80%] | 29% |
| DW | Claude Opus 4.7 | EasyEA | 6.07 | 99% | [97%, 100%] | 50% |
| DW | GPT-4o-mini | NeuSymEA | 4.34 | 62% | [52%, 71%] | 41% |
| DW | GPT-4o | NeuSymEA | 6.39 | 81% | [73%, 88%] | 46% |
| DW | Claude Opus 4.7 | NeuSymEA | 8.21 | 95% | [90%, 99%] | 47% |
| DY | GPT-4o-mini | EasyEA | 3.71 | 53% | [43%, 63%] | 43% |
| DY | GPT-4o | EasyEA | 4.80 | 65% | [56%, 74%] | 53% |
| DY | Claude Opus 4.7 | EasyEA | 6.27 | 83% | [75%, 90%] | 65% |
| DY | GPT-4o-mini | NeuSymEA | 4.34 | 57% | [47%, 66%] | 39% |
| DY | GPT-4o | NeuSymEA | 6.76 | 90% | [84%, 95%] | 48% |
| DY | Claude Opus 4.7 | NeuSymEA | 7.97 | 94% | [89%, 98%] | 48% |
| BIO | GPT-4o-mini | EasyEA | 3.26 | 54% | [44%, 64%] | 32% |
| BIO | GPT-4o | EasyEA | 4.17 | 61% | [51%, 70%] | 35% |
| BIO | Claude Opus 4.7 | EasyEA | 5.88 | 92% | [86%, 97%] | 51% |
Key Findings¶
- Ubiquitous Anchor Bias and Discrimination Inversion: In standard judge configurations displaying candidate decision labels (Prompts I & II), J-ROC-AUC collapses to 0.12–0.53. Judges routinely exhibit inverted discrimination, assigning systematically higher justification scores to incorrect decisions than to correct ones by rationalizing displayed labels.
- Inability of Chain-of-Thought Prompting to Mitigate Bias: Imposing strict 8-step CoT constraints requiring memory suppression and explicit evidence citations (Prompt I) completely fails to remediate anchor bias. In several conditions, CoT provides a flexible narrative space for models to rationalize false premises, worsening anchor adherence.
- Confirmation of the Frontier Model Paradox: Counterfactual experiments confirm that label sensitivity scales monotonically with judge capability. Claude Opus 4.7 exhibits catastrophic label sensitivity (FSR of 83%–99% and MAS exceeding 8.0), whereas GPT-4o-mini demonstrates lower sensitivity (FSR 53%–64%). Stronger reasoning architectures are more prone to confirmation bias when presented with an anchor.
- Architectural Recovery via Label-Free Prediction: Completely removing system labels (Prompt III) restores near-ceiling discrimination performance (J-ROC-AUC 0.931–1.000, Score Gaps +6.0 to +9.1), demonstrating that LLMs maintain robust intrinsic alignment discernment that is artificially suppressed by label exposure.
Highlights & Insights¶
- Formalizing Anchor Bias as a Distinct Failure Mode in Structured NLP: The paper makes a fundamental distinction between open-ended generation biases (verbosity, position) and discrete decision anchoring, formalizing how binary decision labels hijack internal reasoning.
- Counter-Intuitive Frontier Model Paradox: Demonstrates that model scaling alone does not cure sycophantic rationalization—frontier models deploy their superior reasoning capacity to build sophisticated post-hoc justifications for false labels.
- Public Release of MeSH–SNOMED CT 15K Benchmark: Delivers the first large-scale biomedical entity alignment benchmark complete with rich ontological structure, exposing pre-training sparsity vulnerabilities inaccessible via Wikipedia benchmarks.
Limitations & Future Work¶
- Evaluation Under Balanced Class Distribution: The benchmark employs 50/50 class-balanced sampling to isolate discrimination capability, whereas real-world graph alignment involves extreme class imbalance with quadratic candidate spaces.
- Scope of Frontier Model Families: Experiments evaluated three leading closed-source frontier models accessed via OpenRouter; validating whether open-source models (e.g., LLaMA, Qwen) replicate the frontier model paradox remains open.
- Inference Cost of Label-Free Evaluation: Deploying LLMs as independent predictors across all candidate entity pairs incurs higher computational and token expenditure compared to lightweight heuristic filtering.
Related Work & Insights¶
- vs LLM-as-a-Judge Evaluation (Zheng et al., 2023b; Fu et al., 2024): Expands auditing beyond surface-level biases in free-form generation into structured prediction where discrete decisions are evaluated against graph contexts.
- vs Entity Alignment Benchmarks (Sun et al., 2020): Extends evaluation beyond homogeneous DBpedia-Wikidata graphs into challenging cross-schema biomedical ontologies with distinct formal logics.
- vs Shortcut Learning & Robustness (Geirhos et al., 2020; Mirzadeh et al., 2025): Demonstrates that LLM judges rely heavily on lexical entity recognition shortcuts and confirms causal susceptibility through counterfactual intervention.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ [Pioneering identification and formalization of anchor bias and the frontier model paradox in structured evaluation]
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive suite of 153k API calls, causal counterfactuals, perturbation diagnostics, and blinded human validation]
- Writing Quality: ⭐⭐⭐⭐⭐ [Exceptionally coherent exposition, clear conceptual formulations, and rigorous statistical analysis]
- Value: ⭐⭐⭐⭐⭐ [Provides critical actionable guidelines reshaping automated evaluation protocols for knowledge graphs and structured prediction]