Llm debiasing reduces citation bias but lowers judgement clarity
Debiasing as a Measurement Intervention: Calibrated Ties and Resolution Loss in LLM-as-a-Judge Evaluation
Digital LibrariesArtificial Intelligence
Summary
Sometimes AI judges get distracted by how information is shown, like fancy citations, which can bias their decisions. The authors found that asking these judges to ignore such hints can cut down bias but also makes it harder for them to clearly decide which option is better. They built a special test set called TraceJudgeBench to understand this trade-off and showed that some ways of reducing bias turn many decisions into ties, even when there are real differences. The research helps explain how stopping bias in AI judges affects how well their judgments reflect true quality.
What this means in practice
- •For ai system developers: Design AI evaluation protocols that balance bias suppression and judgment clarity using calibrated tie handling and decoupled judging techniques.
- •For automated content reviewers: Improve AI judge prompt engineering to reduce citation-driven bias while maintaining accurate resolution in content quality comparisons.
Authors
Liang Zhao, Yong Wang, Jiangzhe Chen
Abstract
LLM-as-a-judge protocols are commonly debiased by instructing judges to ignore presentation cues such as citation formatting, source labels, and evidence-display style. We show that this intervention can suppress bias while damaging the resolution of the measurement instrument. We introduce TraceJudgeBench, a diagnostic benchmark for auditing citation-like artifacts in RAG and agent-workflow evaluation, covering content-equivalent pairs, citation ablations, correctness conflicts, human-validated soft and moderate quality gaps, prompt-strength ladders, decoupled judging, and a controlled workflow-ranking probe. Across GPT-5.5, Claude Sonnet 4.6, and DeepSeek V4-Flash, stronger anti-citation prompts reduce worse-cited wins from up to 50.5% to 0%; yet some operating points already convert validated moderate-gap decisions into Tie before the strict stress-test endpoint, while correctness-conflict accuracy remains at or above 93.0%. A second, 50-pair FinQA moderate-gap construction reproduces the qualitative frontier, and open-weight Qwen2.5-14B-Instruct-AWQ and Gemma-3-12B-IT runs reproduce the central HotpotQA frontier. TRACE-style decoupling recovers 96.5-100.0% better-plain resolution across the reported settings. Human validation separates three meanings of Tie: correct equivalence Tie, calibrated soft-boundary Tie, and resolution-destroying Tie on validated quality gaps. We frame debiasing as a measurement intervention whose bias suppression, resolution retention, Tie cost, and protocol cost must be reported jointly. The supplementary artifact contains benchmark splits, prompts, raw judge outputs, validation summaries, and analysis.