Research map / Large language models / Language, safety and interpretability
CARE: Confounder-Aware Aggregation for Reliable LLM Evaluation
Jitian Zhao, Changho Shin, Tzu-Heng Huang, Satya Sai Srinath Namburi GNVV, Frederic Sala
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 shared confounders (CARE-SVD) instead of majority vote | LLM-judge aggregation · UltraFeedback | MAE: 0.851 -> 0.623 | Improved |
| model shared confounders (CARE-SVD) instead of simple averaging | LLM-judge aggregation · Review-5K | MAE: 2.274 -> 1.957 | Improved |
| use CARE-Tensor instead of best baseline | LLM-judge aggregation · Summarize | accuracy: 0.718 -> 0.814 | Improved |
| use CARE-SVD on preference data | LLM-judge aggregation · SHP | accuracy: 0.634 -> 0.543 | Hurt |
| use CARE-SVD instead of CARE-Tensor | LLM-judge aggregation · PKU-BETTER | accuracy: 0.779 -> 0.691 | Hurt |
| CARE-Tensor vs MACE on PKU-SAFER | LLM-judge aggregation · PKU-SAFER | accuracy: 0.735 -> 0.731 | No clear effect |
| add programmatic judges via greedy selection with CARE-SVD | CARE-SVD · FeedbackQA | MAE: 15.2% lower than MV | Improved |
| use CARE-SVD under beauty bias injection | LLM-judge aggregation · Beauty bias | MAE shift: 0.506 -> 0.375 | Improved |
| use CARE-SVD under authority bias injection | LLM-judge aggregation · Authority bias | MAE shift: 0.271 -> 0.233 | Improved |
| use CARE-Tensor vs MV on adversarial token attacks | LLM-judge aggregation · Adversarial responses (Thought process:) | false positive rate: 0.674 -> 0.000 | Improved |
| use CARE-Tensor vs WS on comma attack | LLM-judge aggregation · Adversarial response ',' | false positive rate: 0.546 -> 0.000 | Improved |
| replace WS with CARE-SVD (WS on PKU-BETTER is weak) | weak supervision · PKU-BETTER | accuracy: 0.575 vs 0.691 | Mixed |
Closest papers by meaning
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.88)
- Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge (ICLR 2025 · similarity 0.87)
- Limits to scalable evaluation at the frontier: LLM as judge won’t beat twice the data (ICLR 2025 · similarity 0.86)
- Validating LLM-as-a-Judge Systems under Rating Indeterminacy (NeurIPS 2025 · similarity 0.86)
- Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking (ICLR 2025 · similarity 0.86)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.86)
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.
