Related papers: Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation Independence

Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation Independence

URL: http://arxiv.org/abs/2608.25869v1
Date: Wed, 26 Aug 2026 14:41:24 GMT
Title: Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation Independence
Abstract summary: We show that prior scores, even when included only as context metadata, anchor judgments and systematically shift ratings toward their values.<n>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.
Score: 0.7340017786387767
License: http://creativecommons.org/licenses/by/4.0/
Abstract: 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.

Related papers

Beyond Shapley: An Influence-Based Data Auditing Pipeline for LLM Alignment and Evaluation [2.733105115068754]
We introduce a scalable, inference-only data valuation pipeline that approximates the Shapley value without iterative model retraining.<n>We demonstrate the pipeline's efficacy in sanitizing two heavily vetted alignment datasets.
arXiv Detail & Related papers (2026-07-24T03:31:44Z)
Beyond Completion: Probing Cumulative State Tracking to Predict LLM Agent Performance [9.771590610969918]
We introduce Working Memory Fidelity-Active Manipulation (WMF-AM)<n>We evaluate it on 20 open-weight models (0.5B-35B, 13 families) against a released deterministic 10-task agent battery.
arXiv Detail & Related papers (2026-03-28T17:25:11Z)
CodeFuse-CommitEval: Towards Benchmarking LLM's Power on Commit Message and Code Change Inconsistency Detection [8.631593963090985]
Version control relies on commit messages to convey the rationale for code changes, but these messages are often low quality and inconsistent with their diffs-known as message-code inconsistency (MCI)<n>We introduce CODEFUSE-COMMITEVAL, the first benchmark designed for MCI detection using large language models (LLMs)<n>We generate seven types of inconsistent messages through rule-guided mutations of originally consistent commits and apply two-fold validation to verify both positive and negative samples.
arXiv Detail & Related papers (2025-11-25T03:33:57Z)
Reference-Free Rating of LLM Responses via Latent Information [53.463883683503106]
We study the common practice of asking a judge model to assign Likert-scale scores to free-text responses.<n>We then propose and evaluate Latent Judges, which derive scalar ratings from internal model signals.<n>Across a broad suite of pairwise and single-rating benchmarks, latent methods match or surpass standard prompting.
arXiv Detail & Related papers (2025-09-29T12:15:52Z)
TrustJudge: Inconsistencies of LLM-as-a-Judge and How to Alleviate Them [58.04324690859212]
Large Language Models (LLMs) as automated evaluators (LLM-as-a-judge) has revealed critical inconsistencies in current evaluation frameworks.<n>We identify two fundamental types of inconsistencies: Score-Comparison Inconsistency and Pairwise Transitivity Inconsistency.<n>We propose TrustJudge, a probabilistic framework that addresses these limitations through two key innovations.
arXiv Detail & Related papers (2025-09-25T13:04:29Z)
Reducing Biases in Record Matching Through Scores Calibration [1.5530839016602822]
We propose a threshold-independent framework for measuring and reducing score bias.<n>We show that several state-of-the-art matching methods exhibit substantial score bias, even when appearing fair under standard threshold-based metrics.<n>We introduce two post-processing score calibration algorithms. The first, calib, aligns group-wise score distributions using the Wasserstein barycenter, targeting demographic parity.<n>The second, ccalib, conditions on predicted labels to further reduce label-dependent biases, such as equal opportunity.
arXiv Detail & Related papers (2024-11-03T21:01:40Z)
Self-Evaluation Improves Selective Generation in Large Language Models [54.003992911447696]
We reformulate open-ended generation tasks into token-level prediction tasks. We instruct an LLM to self-evaluate its answers. We benchmark a range of scoring methods based on self-evaluation.
arXiv Detail & Related papers (2023-12-14T19:09:22Z)
AMRFact: Enhancing Summarization Factuality Evaluation with AMR-Driven Negative Samples Generation [57.8363998797433]
We propose AMRFact, a framework that generates perturbed summaries using Abstract Meaning Representations (AMRs) Our approach parses factually consistent summaries into AMR graphs and injects controlled factual inconsistencies to create negative examples, allowing for coherent factually inconsistent summaries to be generated with high error-type coverage.
arXiv Detail & Related papers (2023-11-16T02:56:29Z)
Bring Your Own Data! Self-Supervised Evaluation for Large Language Models [52.15056231665816]
We propose a framework for self-supervised evaluation of Large Language Models (LLMs) We demonstrate self-supervised evaluation strategies for measuring closed-book knowledge, toxicity, and long-range context dependence. We find strong correlations between self-supervised and human-supervised evaluations.
arXiv Detail & Related papers (2023-06-23T17:59:09Z)
TRUST-LAPSE: An Explainable and Actionable Mistrust Scoring Framework for Model Monitoring [4.262769931159288]
We propose TRUST-LAPSE, a "mistrust" scoring framework for continuous model monitoring. We assess the trustworthiness of each input sample's model prediction using a sequence of latent-space embeddings. Our latent-space mistrust scores achieve state-of-the-art results with AUROCs of 84.1 (vision), 73.9 (audio), and 77.1 (clinical EEGs)
arXiv Detail & Related papers (2022-07-22T18:32:38Z)
Exploiting Sample Uncertainty for Domain Adaptive Person Re-Identification [137.9939571408506]
We estimate and exploit the credibility of the assigned pseudo-label of each sample to alleviate the influence of noisy labels. Our uncertainty-guided optimization brings significant improvement and achieves the state-of-the-art performance on benchmark datasets.
arXiv Detail & Related papers (2020-12-16T04:09:04Z)

This list is automatically generated from the titles and abstracts of the papers in this site.