CAT-CoT: instruction-tuning LLMs via cognitive appraisal theory-inspired chain-of-thought reasoning to enhance emotional expressivity

Abstract

Despite their conversational proficiency, large language models (LLMs) have a limited potential for true emotional comprehension, generally relying on surface-level indicators such as emotion keywords or sentiment heuristics. This gap becomes critical in domains like mental health support and human-robot interaction, where emotional misalignment can lead to user harm or reduced trust. To address this, we introduce CAT-CoT, a novel CoT framework that integrates Cognitive Appraisal Theory (CAT) with LLM via instruction tuning on a dataset named Appraisal-CoT. We created the Appraisal-CoT dataset, a synthetic corpus of 4,641 empathy-driven dialogues based on the EmpatheticDialogues benchmark and annotated with stepwise appraisal reasoning using GPT-4o-mini. Empirical evaluation demonstrates that our instruction-tuned models significantly outperform the untuned baselines on EmoBench, achieving the highest gains of 6.5% in Emotion Understanding (EU) and 4.0% in Emotion Application (EA) with Qwen3-4B-ACoT, compared to the base model. Also, the small models score matching or exceeding the performance of larger models (e.g., Qwen3 4B–ACoT rivals 8B variants). These findings demonstrate that employing psychologically grounded reasoning techniques can significantly enhance emotional expressivity in LLMs, marking a step toward safer, more empathetic AI systems.