Research map / Large language models / Language, safety and interpretability
Diagnosing the Reliability of LLM-as-a-Judge via Item Response Theory
Junhyuk Choi, Sohhyung Park, Chanhee Cho, Hyeonchu Park, Bugeun Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add detailed evaluation instructions with rubric definitions (ablation) | LLM judge prompts · TopicalChat | prompt consistency CV: GPT-4o Naturalness 0.28 -> 0.10 | Improved |
| add chain-of-thought on top of detailed prompt (ablation) | LLM judge prompts · TopicalChat Naturalness | prompt consistency CV: GPT-4o 0.10 -> 0.01 | Improved |
| add chain-of-thought on top of detailed prompt (ablation) | LLM judge prompts · TopicalChat | marginal reliability rho | No clear effect |
| use 5-point instead of 3-point rating scale (ablation) | LLM judges · TopicalChat Naturalness/Coherence/Engagingness | marginal reliability rho: 0.91-0.95 (Naturalness) | Improved |
| use 7-point instead of 5-point rating scale (ablation) | LLM judges · TopicalChat | marginal reliability rho | No clear effect |
| use larger model (Qwen3-235B vs Qwen3-30B) (ablation) | LLM judges · SummEval Relevance | prompt consistency CV: 0.17 -> 0.09 | Improved |
| use larger model (GPT-4o vs GPT-4o-mini) (ablation) | LLM judges · HelpSteer-2 Helpfulness | marginal reliability rho: 0.67 -> 0.84 | Improved |
| use larger model (Qwen3-VL-235B vs 30B, GPT-4o vs mini) (ablation) | VLM judges · VIEScore | CV and rho | No clear effect |
| evaluate vision-language judge (VIEScore) instead of text judges | LLM judges · VIEScore vs NLP | prompt consistency CV: NLP CV < 0.30 vs VIEScore 0.16 to 1.32 | Hurt |
| judge perceptual quality (PQ) with VLMs | VLM judges · VIEScore PQ | discrimination breadth ratio: theta ratio 2.03-4.40 | Hurt |
Closest papers by meaning
- Validating LLM-as-a-Judge Systems under Rating Indeterminacy (NeurIPS 2025 · similarity 0.86)
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.86)
- Evaluating the Evaluator: Measuring LLMs’ Adherence to Task Evaluation Instructions (AAAI 2025 · similarity 0.86)
- Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge (ICLR 2025 · similarity 0.85)
- Evaluating Judges as Evaluators: The JETTS Benchmark of LLM-as-Judges as Test-Time Scaling Evaluators (ICML 2025 · similarity 0.84)
- Can We Trust Item Response Theory for AI Evaluation? (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
