Research map / Large language models / Language, safety and interpretability
Dependence-Aware Label Aggregation for LLM-as-a-Judge via Ising Models
Krishna Balasubramanian, Aleksandr Podkopaev, Shiva Kasiviswanathan
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 |
|---|---|---|---|
| model judge dependence with Ising models instead of conditional independence (weighted majority vote) | LLM-as-a-judge label aggregation · Relevance (WikiQA) | Test accuracy: 0.820 -> 0.912 | Improved |
| model judge dependence with Ising models instead of CI weighted majority vote | LLM-as-a-judge label aggregation · Toxicity (Jigsaw) | Test accuracy: 0.694 -> 0.792 | Improved |
| model judge dependence with Ising models instead of CI weighted majority vote | LLM-as-a-judge label aggregation · Summarization (CNN/DailyMail) | Test accuracy: 0.737 -> 0.801 | Improved |
| use class-dependent couplings instead of class-independent couplings (ablation) | Ising aggregator · Relevance (WikiQA) | Test accuracy: 0.899 -> 0.912 | Improved |
| use class-dependent couplings instead of class-independent couplings (ablation) | Ising aggregator · Toxicity (Jigsaw) | Test accuracy: 0.780 -> 0.792 | Improved |
| use class-dependent couplings instead of class-independent couplings (ablation) | Ising aggregator · Summarization (CNN/DailyMail) | Test accuracy: 0.806 -> 0.801 | No clear effect |
| use weighted instead of uniform majority vote (ablation) | CI aggregator · Summarization (CNN/DailyMail) | Test accuracy: 0.561 -> 0.737 | Improved |
| use weighted instead of uniform majority vote (ablation) | CI aggregator · Toxicity (Jigsaw) | Test accuracy: 0.695 -> 0.694 | No clear effect |
| increase number of judges or training samples (ablation) | Ising predictors vs CI-WMV · all three datasets | Test accuracy | Improved |
| ignore dependence (CI predictor) under class-conditional Curie-Weiss model | CI predictor · theory, K to infinity | risk: Bayes risk -> 0 vs CI risk min{pi, 1-pi} | Hurt |
Closest papers by meaning
- Limits to scalable evaluation at the frontier: LLM as judge won’t beat twice the data (ICLR 2025 · similarity 0.87)
- Distribution-Calibrated Inference Time Compute for Thinking LLM-as-a-Judge (ICML 2026 · similarity 0.87)
- Validating LLM-as-a-Judge Systems under Rating Indeterminacy (NeurIPS 2025 · similarity 0.84)
- Operads for compositional reasoning in LLMs (NeurIPS 2026 · similarity 0.84)
- Investigating Non-Transitivity in LLM-as-a-Judge (ICML 2025 · similarity 0.84)
- Don’t Label Twice: Quantity Beats Quality when Comparing Binary Classifiers on a Budget (ICML 2024 · 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.
