CARE

Researchers propose CARE (Contrastive Anchor-based Rubric Evolution), a method that uses anchor responses from frontier models to dynamically evolve rubrics during LLM post-training, addressing reward…

Researchers introduced CARE (Community-Aware Reaction Evaluation), a human-AI collaborative framework that benchmarks large language models against authentic community responses to real-world news by …