CARE: Cognitive-reasoning Augmented Reinforcement for Emotional Support Conversation

TL;DR

CARE tackles the gap in cognitive reasoning for Emotional Support Conversation by augmenting responses with explicit reasoning chains drawn from the original ESC dataset rather than relying on large synthetic corpora. It defines four reasoning nodes—Context, Cognition, Emotion, and Support Plan—and uses reinforcement learning with multi-dimensional rewards to strengthen both the reasoning process and final support quality. The approach yields higher automatic metrics and favorable human evaluations on ESConv compared to strong baselines, with the SFT-RL variant achieving the best results. These findings suggest that structured cognitive reasoning can produce more logical, empathetic, and human-like ESC systems with limited data augmentation. The work implies practical benefits for deploying emotionally supportive AI in real-world settings.

Abstract

Emotional Support Conversation (ESC) plays a vital role in alleviating psychological stress and providing emotional value through dialogue. While recent studies have largely focused on data augmentation and synthetic corpus construction, they often overlook the deeper cognitive reasoning processes that underpin effective emotional support. To address this gap, we propose \textbf{CARE}, a novel framework that strengthens reasoning in ESC without relying on large-scale synthetic data. CARE leverages the original ESC training set to guide models in generating logically coherent and supportive responses, thereby explicitly enhancing cognitive reasoning. Building on this foundation, we further employ reinforcement learning to refine and reinforce the reasoning process. Experimental results demonstrate that CARE significantly improves both the logical soundness and supportive quality of responses, advancing the development of empathetic, cognitively robust, and human-like emotional support systems.

Figures (3)

• Figure 1: Overview of CARE framework.
• Figure 2: Human evaluation results comparing CARE with three baselines, showing CARE’s win, tie, and loss percentages.
• Figure 3: Case study: CARE vs. baselines in a job-loss scenario.