REVIEW 30 cited by
Large Language Models are Inconsistent and Biased Evaluators
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Large Language Models are Inconsistent and Biased Evaluators
read the original abstract
The zero-shot capability of Large Language Models (LLMs) has enabled highly flexible, reference-free metrics for various tasks, making LLM evaluators common tools in NLP. However, the robustness of these LLM evaluators remains relatively understudied; existing work mainly pursued optimal performance in terms of correlating LLM scores with human expert scores. In this paper, we conduct a series of analyses using the SummEval dataset and confirm that LLMs are biased evaluators as they: (1) exhibit familiarity bias-a preference for text with lower perplexity, (2) show skewed and biased distributions of ratings, and (3) experience anchoring effects for multi-attribute judgments. We also found that LLMs are inconsistent evaluators, showing low "inter-sample" agreement and sensitivity to prompt differences that are insignificant to human understanding of text quality. Furthermore, we share recipes for configuring LLM evaluators to mitigate these limitations. Experimental results on the RoSE dataset demonstrate improvements over the state-of-the-art LLM evaluators.
Forward citations
Cited by 30 Pith papers
-
Refused in Chat, Written in Code: Workflow-Level Jailbreak Construction in IDE Coding Agents
Four Copilot backends refuse almost all 204 harmful prompts in chat and simple framings, yet produce 816/816 unsafe teaching-shot completions under a multi-turn IDE evaluation-pipeline workflow.
-
Refused in Chat, Written in Code: Workflow-Level Jailbreak Construction in IDE Coding Agents
Four Copilot backends refuse almost all harmful prompts in chat or simple framings, yet produce 816/816 unsafe teaching-shot completions under a multi-turn IDE evaluation-pipeline workflow.
-
Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior
Post-training recipe shapes hedging behavior in interactive multi-LLM systems more than model family, with within-family variation exceeding cross-family gaps on a 940k-chain corpus.
-
OR-Space: A Full-Lifecycle Workspace Benchmark for Industrial Optimization Agents
OR-Space is a benchmark for LLM agents performing full-lifecycle optimization tasks across Build, Revise, and Explain modes in executable multi-artifact workspaces.
-
GRASP: Deterministic argument ranking in interaction graphs
GRASP aggregates stable local LLM interaction judgments into global argument rankings via a convergent attack-defense propagation operator on interaction graphs, yielding higher reproducibility than holistic judging a...
-
ProactBench: Beyond What The User Asked For
ProactBench measures LLM conversational proactivity in three phases using 198 multi-agent dialogues and finds recovery behavior hard to predict from existing benchmarks.
-
TREK: A Travel Reasoning and Evaluation Kit for LLM Agents in Complex Trip Planning
On 800 jointly constrained trip tasks with a deterministic scorer and achievable gold, the best of 15 LLM agents fully solves only 46.2% of feasible plans, with unstated persona needs as the universal bottleneck.
-
Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI
A post-hoc audit framework finds that many agent benchmark scores are inflated by protocol exposures (public answers, readable hidden state, generator regularities, feedback, or scoring flaws), with measured inflation...
-
Persuasion Attacks Can Decrease Effectiveness of CoT Monitoring
Adversarial agents can exploit visible chain-of-thought reasoning to persuade monitor LLMs to approve policy-violating actions, but cross-family fact-checking reduces approval rates by up to 45%.
-
The Blind Curator: How a Biased Judge Silently Disables Skill Retirement in Self-Evolving Agents
False-pass bias in LLM judges disables contribution-based skill retirement past a sharp threshold, while symmetric noise does not, and the failure is often silent in aggregate metrics.
-
Necessary but Not Sufficient: Temperature Control and Reproducibility in LLM-as-Judge Safety Evaluations
Even with temperature pinned at 0 and greedy decoding, 1-2 of 7 borderline safety items flip pass/fail verdicts across runs on real LLM providers, and some models deprecate temperature control entirely.
-
Reliability without Validity: A Systematic, Large-Scale Evaluation of LLM-as-a-Judge Models Across Agreement, Consistency, and Bias
Large-scale study of 21 LLM-as-a-Judge models shows exact-match agreement overstates reliability, rankings shift across benchmarks, and high consistency can mask position bias.
-
From `May' to `Is': Certainty Distortion in Language Model Rewriting
LMs systematically inflate expressed certainty during rewriting, affecting up to 75% of outputs with a 1.5-2x bias toward increasing rather than decreasing certainty, and the effect compounds over iterations.
-
Auditing Stance Asymmetry in Generative Explanations
Introduces Symmetry Decomposition Evaluation (SDE) to audit stable stance asymmetries in generative explanations using paired situations, role rewrites, and evidence controls on a 32-family prototype suite.
-
LP-Eval: Rubric and Dataset for Measuring the Quality of Legal Proposition Generation
LP-Eval is a new expert-co-designed rubric and annotated dataset showing that LLMs mostly produce well-formed legal propositions from EU court decisions, with higher expert-rated quality for established cases and impr...
-
Coordinates of Capability: A Unified MTMM-Geometric Framework for LLM Evaluation
A new MTMM-geometric framework unifies LLM evaluation metrics into three latent dimensions to separate method variance from true capabilities.
-
How Compliant Are GitHub Actions Workflows? A Checklist-Based Study with LLM-Assisted Auditing
GitHub Actions workflows achieve only 28% overall compliance with best practices, with LLMs enabling an 81% reduction in verification effort via hybrid adjudication but still requiring expert oversight for security judgments.
-
Self-Preference Bias in LLM-as-a-Judge
LLMs judge their own outputs higher because they assign better scores to lower-perplexity text, even when the text is not self-generated.
-
Pramana: A Protocol-Layer Treatment of Claim Verification in Autonomous Agent Networks
Pramana defines a typed ClaimAttestation protocol with four variants and verify operations, specifies its lifecycle in TLA+, model-checks it with TLC, and provides a tested Python implementation for auditable agent claims.
-
Coordinates of Capability: A Unified MTMM-Geometric Framework for LLM Evaluation
A systematization of knowledge unifies nine LLM metrics into three orthogonal latent dimensions via an MTMM-geometric framework to improve construct validity in evaluation.
-
STELLAR-E: a Synthetic, Tailored, End-to-end LLM Application Rigorous Evaluator
STELLAR-E modifies the TGRT Self-Instruct framework to produce tailored synthetic LLM evaluation datasets that score an average 5.7% higher on LLM-as-a-judge metrics than existing language-specific benchmarks.
-
Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines
Gemini 2.5 Flash with a Combined Budget debiasing strategy achieves 71.0% judge agreement at ~$0.001/evaluation, outperforming frontier models at 15x lower cost.
-
Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines
Style bias dominates LLM-as-a-Judge systems far more than position bias, with debiasing strategies providing model-dependent gains and public tools released for replication.
-
On Cost-Effective LLM-as-a-Judge Improvement Techniques
Ensemble scoring plus task-specific criteria injection raises LLM judge accuracy to 85.8 percent on RewardBench 2, a 13.5-point gain over baseline, with small models gaining the most.
-
On Cost-Effective LLM-as-a-Judge Improvement Techniques
Ensemble scoring plus task-specific criteria injection lifts LLM judges to 85.8% on RewardBench 2 (+13.5pp), dominating calibration and escalation on the cost–accuracy frontier.
-
CODE-GEN: A Human-in-the-Loop RAG-Based Agentic AI System for Multiple-Choice Question Generation
A tool-augmented dual-agent RAG system generates and validates coding MCQs, reaching 79.9–98.6% human-validated success and showing AI handles computational checks while humans remain essential for pedagogical depth.
-
Verify-Gated Completion as Admission Control in a Governed Multi-Agent Runtime: A Bounded Architecture Case Study
In a bounded multi-agent runtime case study, a read-only verify gate achieved 99.5% success on known-outcome invoked events and supported inspectable, fail-closed completion decisions under the observed conditions.
-
Verify-Gated Completion as Admission Control in a Governed Multi-Agent Runtime: A Bounded Architecture Case Study
In a bounded multi-agent runtime case study, verify-gated completion produced 99.5% success on invoked verification events with packetized records, supporting only a narrow claim of inspectable and fail-closed decisio...
-
LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods
A survey that organizes LLMs-as-judges research into functionality, methodology, applications, meta-evaluation, and limitations.
-
Position: Multimodal Large Language Models Can Significantly Advance Scientific Reasoning
Position paper claims multimodal LLMs can significantly advance scientific reasoning and proposes a four-stage roadmap plus challenges and suggestions.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.