EiCAP: Beyond Fluency, Probing and Improving Emotional Intelligence in LLMs via Psychologically Grounded Multi-Turn Dialogue

The paper introduces EiCAP, a framework for assessing and enhancing Emotional Intelligence (EI) in Large Language Models (LLMs) through a six-layer EI taxonomy. It presents two resources: EiCAP-Bench, a benchmark with 3,174 probes across 24 EI subcategories, and EiCAP-SFT, a 152,820-dialogue corpus for fine-tuning. Notably, applying EI-grounded LoRA to the Qwen-2.5-7B-Base model, utilizing only 0.8% of parameters, resulted in a significant performance improvement, achieving a macro score of 75.33%, highlighting the importance of targeted EI training over generic conversational fine-tuning.