CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels

Researchers introduced CARE, a parameter-efficient method for learning from long-tailed datasets with noisy labels.

CARE adaptively tightens agreement for rare classes and relaxes it for frequent classes by combining noisy labels, VLM text embeddings, and visual features.

Across synthetic and real benchmarks, it improved accuracy by up to 3.0% over state-of-the-art methods, with especially strong gains on underrepresented classes.