AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
model shared confounders (CARE-SVD) instead of majority voteLLM-judge aggregation · UltraFeedbackMAE: 0.851 -> 0.623Improved
model shared confounders (CARE-SVD) instead of simple averagingLLM-judge aggregation · Review-5KMAE: 2.274 -> 1.957Improved
use CARE-Tensor instead of best baselineLLM-judge aggregation · Summarizeaccuracy: 0.718 -> 0.814Improved
use CARE-SVD on preference dataLLM-judge aggregation · SHPaccuracy: 0.634 -> 0.543Hurt
use CARE-SVD instead of CARE-TensorLLM-judge aggregation · PKU-BETTERaccuracy: 0.779 -> 0.691Hurt
CARE-Tensor vs MACE on PKU-SAFERLLM-judge aggregation · PKU-SAFERaccuracy: 0.735 -> 0.731No clear effect
add programmatic judges via greedy selection with CARE-SVDCARE-SVD · FeedbackQAMAE: 15.2% lower than MVImproved
use CARE-SVD under beauty bias injectionLLM-judge aggregation · Beauty biasMAE shift: 0.506 -> 0.375Improved
use CARE-SVD under authority bias injectionLLM-judge aggregation · Authority biasMAE shift: 0.271 -> 0.233Improved
use CARE-Tensor vs MV on adversarial token attacksLLM-judge aggregation · Adversarial responses (Thought process:)false positive rate: 0.674 -> 0.000Improved
use CARE-Tensor vs WS on comma attackLLM-judge aggregation · Adversarial response ','false positive rate: 0.546 -> 0.000Improved
replace WS with CARE-SVD (WS on PKU-BETTER is weak)weak supervision · PKU-BETTERaccuracy: 0.575 vs 0.691Mixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy