Source-linked AI summary
Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation Independence
Ante Kapetanovic, Kemal Altwlkany, Andro Mercep, Tomislav Duricic, Emanuel Lacic
TL;DR
LLM-as-a-Judge systems are commonly treated as independent evaluators even when prior-evaluation metadata remains in context. This paper tests that assumption across controlled numerical and categorical evaluations, finding systematic effects from anchored metadata and limited mitigation in the tested settings.
Problem
The paper asks whether prior-evaluation metadata compromises the assumed independence of LLM-as-a-Judge judgments in production evaluation workflows.
Method
The study compares no metadata, revision framing, and anchored metadata across eight models, 20 fixed texts, task categories, token-probability probes, and categorical industry data.
Results
Seven of eight evaluated models show 95% task-stratified bootstrap intervals below zero for the total anchored-metadata effect, with categorical effects also observed.
Takeaways & Limitations
LLM evaluation should measure total condition effects and validate mitigation strategies for each intended model, task, domain, and workflow.
Takeaways & Limitations
Conclusions are bounded by 20 fixed benchmark texts, anchors in [0, 3.99), one 441-sample industry domain, selected token probes, and an altered C2 metadata template.
Abstract
from arXiv · showhide
Large language models (LLMs) increasingly assess generated content, giving rise to the LLM-as-a-Judge paradigm. These systems now score outputs, filter content, and gate iterative refinement in production pipelines, where each judgment is often assumed to be independent of earlier evaluations. We test this assumption using three prompt conditions: no metadata, revision framing, and anchored metadata containing revision, attempt, and prior-score fields. We show that prior scores, even when included only as context metadata, anchor judgments and systematically shift ratings toward their values. Across 192,000 attempted evaluations (185,271 successful), seven out of the eight evaluated models have 95% task-stratified bootstrap intervals below zero for the total anchored-metadata effect on 20 fixed texts. Cohen's $d$, a standardized measure of the difference between score distributions, reaches an absolute value of 0.71. Token-level analysis of selected model-task probes suggests a threshold-like response pattern: introducing anchored metadata produces a marked redistribution of output-score probabilities, while changing the anchor value within the tested below-threshold range produces comparatively little additional variation. On categorical industry data with human-labeled ground truth, anchored metadata blocks 48% of error corrections and flips 10.18% of correct judgments toward an assigned wrong label, demonstrating the bias extends beyond numerical scoring to categorical decisions. Neither Chain-of-Thought nor a metadata-disregard warning reduces the total effect, although the warning improves the paired accuracy effect relative to baseline in the industry experiment. Reliable LLM evaluation demands careful context engineering rather than an assumption of impartiality. Effective mitigation must be validated for the intended model and task or domain.
1 Introduction
LLM-as-a-Judge systems are increasingly used to evaluate generated content in production workflows that implicitly assume each judgment is independent of prior evaluations. This paper investigates whether revision framing and anchored prior-evaluation metadata affect those judgments.
- Motivation: LLM-as-a-Judge systems automate or supplement human evaluation of generated content across benchmarking and production workflows.Applications include preference learning, content moderation, iterative-refinement gates, and multi-agent arbitration.
- Motivation: Prior scores can persist in context across revisions, despite the assumption that subsequent judgments remain independent.The metadata may appear as revision tags, upstream labels, or previous self-evaluations.
- Background: Human anchoring describes numerical estimates being pulled toward an initially observed value, while framing changes choices through equivalent gain-versus-loss descriptions.The paper uses these cognitive-bias findings as background for testing analogous effects in LLM judges.
- Research questions: The study asks how revision framing and complete anchored metadata affect scores, whether responses show threshold-like anchor sensitivity, and which mitigations reduce bias.These questions cover numerical judgments, response patterns, and intervention effectiveness.
- Research questions: The experimental design contrasts no metadata, revision framing, and anchored metadata containing revision, attempt, and prior-score fields.The anchored condition is evaluated against both the baseline and the framing condition.
2 Related Work
Prior work documents cognitive and evaluator-specific biases in LLMs, but anchoring within rubric-based LLM-as-a-Judge scoring remains comparatively understudied. This paper extends the question to below-threshold anchors and production-oriented categorical validation.
- Cognitive and evaluator biases: LLMs exhibit functional analogs of human cognitive biases, while awareness-based prompting has shown mitigation promise for some general cognitive biases.Related findings include framing effects and surveyed evidence across multiple bias types.
- Anchoring in LLMs: Prior anchoring studies report incomplete mitigation from Chain-of-Thought or ignore instructions and identify shallow-layer mechanisms in some models.Other work also finds systematic shifts from source attribution alone.
- Cognitive and evaluator biases: LLM evaluators show position, verbosity, self-preference, consistency, authority, beauty, preference-leakage, and shortcut biases.These findings motivate broader scrutiny of judge behavior beyond score accuracy alone.
- Mitigation literature: Proposed mitigation approaches include multi-reviewer systems, human-preference alignment, calibration, and prompt-level awareness interventions.The literature offers several strategies, but their effectiveness is bias- and setting-dependent.
- Research gap: Anchoring inside LLM-as-a-Judge evaluation remains understudied, especially for rubric-based scoring, below-threshold anchors, and production data.This paper addresses those gaps with fixed-scale rubric judgments, targeted token-probability probes, and industry validation.
3 Methods
The study compares three metadata conditions in fixed-scale rubric grading and estimates anchored-metadata effects across models, tasks, temperatures, and task-aware bootstrap samples. It also defines the evaluation protocol and statistical estimands used for inference.
- Experimental design: The judge receives one candidate answer and assigns a numerical score on a fixed 0–5 rubric with 4.0 as the acceptance threshold.The rubric divides scores into six interpretive bands from unacceptable to excellent.
- Experimental design: C0 omits contextual metadata, C1 adds revision framing, and C2 adds revision framing, an attempt index, and a prior score.C2 prior scores are sampled below the acceptance threshold, while answer texts remain fixed across conditions.
- Estimands: The primary estimand is ΔC2−C0 = μC2 − μC0, while ΔC2−C1 is interpreted as incremental beyond revision framing rather than as the isolated numerical-score effect.C2 also changes the metadata block and adds an attempt field, preventing that contrast from isolating the prior score alone.
- Tasks and models: The evaluation covers eight models and 20 fixed tasks spanning summarization, code review, creative writing, and factual question answering.Each category contains five tasks, and each task uses one screened answer text across conditions and models.
- Evaluation protocol: Across 192,000 attempted calls, 185,271 responses were valid for score analysis, with acceptance derived as score ≥4.0.The protocol used 100 calls for every model–task–condition–temperature combination, and all 1,920 cells retained valid responses.
- Statistical analysis: Condition means are contrasted within model–task–temperature cells, then averaged across temperatures and equally across 20 tasks with 10,000 task-stratified bootstraps.Tasks, rather than repeated decodings, are the inferential resampling units; Cohen’s d is descriptive and computed on complete triplets.
4 Results
Across the fixed eight-model evaluation, anchored metadata generally lowers scores, with effects varying by model, task category, and temperature. These score changes can produce substantial acceptance-rate shifts, especially near the decision threshold.
- 4.1 Anchoring Bias Across Models: Seven of eight models have 95% task-bootstrap intervals below zero for the total anchored-metadata effect ΔC2−C0.
- 4.1 Anchoring Bias Across Models: Cohen's d reaches −0.71 for Llama-3.2-3B and −0.67 for GPT-4.1, while Claude-4.5-Sonnet has the largest absolute score shift at −0.706.
- 4.2 Acceptance-Rate Effects: C2 lowers acceptance by 22.29 percentage points for Claude-4.5-Sonnet and 14.40 percentage points for GPT-4.1.Estimated changes are smaller or positive for some models, and uncertainty does not distinguish the Llama-3.3-70B and Llama-3.1-8B changes from baseline.
- 4.3 Category and Model Heterogeneity: Mean ΔC2−C0 is −0.047 for code review, −0.282 for creative writing, −0.338 for factual QA, and −0.322 for summarization.Code review is weakest and most heterogeneous, and the pattern is not uniform across every model–category combination.
- 4.4 Task-Aware Joint Analysis: The fixed-model-set joint analysis finds the three non-code-review categories more negative than code review, while the code-review intercept interval spans zero.
- 4.4 Task-Aware Joint Analysis: Temperature contrasts against T=0 are small and their intervals include zero, while descriptive mean effects remain negative at every tested temperature.The reported contrasts are +0.011, +0.061, and +0.009 for T=0.5, 1, and 2, respectively.
5 Mechanism Analysis
Targeted token-level probes examine whether anchored metadata changes responses smoothly with anchor magnitude or instead produces a threshold-like redistribution. The clearest pattern appears for GPT-4.1, partially replicates in Llama-3.2-3B, and remains non-universal across the tested probes.
- 5.1 Targeted GPT-4.1 Probe: GPT-4.1's selected creative-writing probe shifts from a mean score of 5.00 in C0 to 4.358 in C2, a ΔC2−C0 of −0.642.The probe used 100 calls per condition at each of four anchor values from 1.0 to 3.9.
- 5.1 Targeted GPT-4.1 Probe: Score-digit probability mass moves from “5” to “4” under C2, while within-C2 slopes over anchor value are approximately zero.This combines a marked C0-to-C2 redistribution with little additional variation across the tested anchor values.
- 5.2 Cross-Model Replication: The selected token-level evidence does not identify which prompt component triggers the redistribution or establish an internal mechanism.A matched template containing an irrelevant number was not used.
- 5.3 Ceiling Robustness: Fifty of 60 model–task effects are negative away from the ceiling, with mean ΔC2−C0 of −0.227.The adjusted association between distance to ceiling and absolute effect magnitude is near zero, so ceiling proximity is not sufficient to explain the shifts.
- 5.2 Cross-Model Replication: The threshold-like response pattern is clearest for GPT-4.1 and partially replicates in Llama-3.2-3B, whose mean score falls by 0.120.Llama-3.1-8B changes by −0.018 with unchanged modal probability, while Llama-3.3-70B changes by +0.016 without token redistribution.
6 Mitigation Strategies
The study tests Chain-of-Thought and an explicit metadata-disregard warning as mitigations for anchored judgments on one GPT-4.1 summarization task. Neither intervention reduces the primary total effect, although the warning flattens the within-anchor slope.
- Experimental design: 100 C0 and 100 C2 calls were run at each of four prior scores for every strategy, with all 1,500 calls valid.The attempt index was fixed at 2, and C2 averages equally over the four anchor values.
- Total effect: CoT shifts Δtotal by −0.6327, a 47.7% worsening in absolute magnitude relative to baseline’s −0.4285.The primary metric is the strategy-specific total effect, Δtotal = μC2 − μC0.
- Total effect: The warning shifts Δtotal by −0.4571, a 6.7% worsening in absolute magnitude, so neither intervention reduces the primary effect.The warning does flatten the secondary within-C2 slope, but slope reduction does not establish total bias reduction.
- Dose-response diagnostics: The warning reduces the within-C2 slope from 0.2674 to −0.0066, whereas CoT increases both the total shift and the slope.These slopes are secondary diagnostics rather than measures of the total anchored-metadata effect.
- Scope: The findings are conditional on one GPT-4.1 summarization pair; broader models, tasks, and interventions remain untested.The paper therefore does not support a general numerical-mitigation recommendation.
7 Industry Validation
The paper validates anchored-metadata effects on categorical industry judgments using human-verified labels and paired comparisons. Anchored metadata blocks corrections of prior errors and induces wrong-label shifts, while mitigation results differ across strategies.
- Data and design: 441 proprietary messaging-campaign samples carry human labels in three classes: compliant, drifting, or shaft.The original Llama-3-8B classifier achieved 49.66% accuracy; only aggregate group compositions are disclosed.
- Data and design: Group A contains 222 original classifier errors, while Group B contains 219 original correct classifications evaluated with fixed sample-specific wrong anchors.GPT-4.1 was evaluated under paired C0 and C2 outputs for Group B.
- Error correction: 10.36 percentage points: Group A correction falls from 21.62% in C0 to 11.26% in C2, blocking 47.9% of baseline corrections.Because C2 changes metadata beyond C1, the C2–C1 contrast does not isolate the prior label.
- Induced errors: −7.76 percentage points: Group B accuracy falls from 76.26% in C0 to 68.49% in C2, while 10.18% of C0-correct cases become the assigned wrong label.Anchored metadata therefore both blocks corrections and induces paired errors.
- Mitigation effectiveness: Warning improves the paired accuracy effect relative to baseline by 7.76 percentage points, whereas CoT’s difference from baseline is uncertain.Warning also reduces the increase in wrong-anchor agreement by 6.39 percentage points in magnitude.
- Scope: The mitigation conclusions are restricted to the tested strategies, GPT-4.1, categorical task, and industry domain.The results do not establish a universal mitigation ranking or a direct warning-versus-CoT comparison.
8 Discussion and Conclusion
The findings indicate that prior-evaluation metadata can compromise LLM-judge independence across numerical and categorical decisions, while tested mitigations do not support a general deployment recommendation. Conclusions remain bounded by the benchmark, anchor range, selected probes, and single industry domain.
- Discussion and Conclusion: Seven of eight fixed LLM judges had 95% task-bootstrap intervals below zero for the total anchored-metadata effect across 192,000 attempted evaluations.185,271 responses were valid; supported absolute Cohen’s d values ranged from 0.09 to 0.71, while Llama-3.1-8B’s interval crossed zero.
- Discussion and Conclusion: Token-probability probes most clearly showed a threshold-like response for GPT-4.1, with partial replication in Llama-3.2-3B and little or no redistribution in larger Llama probes.The evidence does not establish a universal threshold mechanism, and ceiling proximity was not sufficient to explain downward shifts.
- Discussion and Conclusion: 47.9% lower correction rates and 10.18% wrong-label flips in industry validation show that anchored metadata affects categorical decisions as well as rubric scores.In the paired Group B experiment, baseline accuracy fell by 7.76 pp and wrong-anchor agreement rose by 7.31 pp.
- Discussion and Conclusion: Neither Chain-of-Thought nor a metadata-disregard warning reduced the total numerical effect, although the warning improved the paired accuracy effect in the tested industry domain.No tested prompt strategy supports a general deployment recommendation; excluding experimentally irrelevant metadata remains the direct safeguard when workflow design permits it.
- Discussion and Conclusion: The conclusions are limited by one screened answer for each of 20 tasks, anchors restricted to [0, 3.99), selected token probes, and one 441-sample industry domain.Broader tasks, answer sets, anchor regimes, matched metadata templates, and real-world corpora are needed to assess transfer and component-specific effects.
- Discussion and Conclusion: The findings concern independence under the tested information state, not models knowingly following random values presented as genuine prior evaluations.The randomized anchors were exogenous to answer quality.
- Discussion and Conclusion: Reliable deployment requires task-aware measurement, cautious interpretation of prompt-field effects, and mitigation validation for each intended model, task, and workflow.Prior-evaluation metadata can compromise independence in both rubric scores and categorical decisions.
Generative AI Tools Use Disclosure
The authors disclose using generative AI tools for code implementation, data analysis, editing, grammar checking, and related-work identification, with all assisted outputs reviewed and approved.
- Generative AI Tools Use Disclosure: Claude Code, ChatGPT, and Google Scholar Labs assisted with implementation, analysis, editing, grammar checking, and related-work identification.The authors state that they reviewed and approved all AI-assisted outputs and take responsibility for the publication.