CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning
The article introduces CAREBench, a new benchmark designed to evaluate the emotion understanding capabilities of large language models (LLMs). It highlights the limitations of existing evaluation methods and proposes a process-level evaluation framework based on cognitive appraisal reasoning. The findings suggest that while some LLMs perform well in certain tasks, they struggle with understanding human emotional complexity.
- ▪CAREBench is the first benchmark with complete inferential chain annotations for evaluating LLMs' emotion understanding.
- ▪The study reveals that stronger models can match or exceed human performance in specific tasks but fall short in appraisal reasoning and positive emotion recognition.
- ▪Current LLMs have not fully internalized the cognitive mechanisms necessary to capture human emotional diversity.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.17176 (cs) [Submitted on 16 May 2026] Title:CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning Authors:Zhaoyue Sun, Hainiu Xu, Andero Uusberg, James J. Gross, Petr Slovak, Yulan He View a PDF of the paper titled CAREBench: Evaluating LLMs' Emotion Understanding by Assessing Cognitive Appraisal Reasoning, by Zhaoyue Sun and 4 other authors View PDF HTML (experimental) Abstract:Emotion understanding is a core capability for LLMs to interact effectively with humans, yet existing evaluation paradigms rely on discrete emotion label prediction and fail to capture the cognitive processes underlying emotion generation.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
