FdCogChain: Self-Feedback Refinement of Cognitive Chain-of-Thought Reasoning with LLMs for Emotional Support Conversations

Emotional support conversations can reduce mental stress and provide social benefits. Cognitive chain-of-thought (CogChain) reasoning helps the supporter infer the help-seeker’s mental state, but existing methods show limited performance in emotion recognition and reasoning depth. This study proposes FdCogChain, a multi-level self-feedback refinement framework based on large language models. The framework automatically refines the initial CogChain into a cognitive model chain-of-thought (CogChain-M) by identifying its shortcomings and generating improved alternatives. CogChain-M enhances four components: it introduces fine-grained emotion modeling, enriches context representation, focuses thoughts on automatic thought and core beliefs, and shifts behavior descriptions toward future behavior prediction. CogChain-M uses a chain structure to analyze the help-seeker’s utterance, context, emotion, thought, and behavior, then selects appropriate support strategies. To improve the efficiency of automatically refining initial CogChains into CogChain-M, this study designs an in-context learning pipeline based on large language models, enabling efficient generation of CogChain-M for any given dialogue and its corresponding CogChain. A new dataset is constructed based on the ESChain dataset, and a supporter model equipped with cognitive reasoning capabilities is trained. Experimental results demonstrate that FdCogChain significantly improves the emotional support performance of the model on both in-domain and out-of-domain tasks, highlighting its effectiveness and generalization ability.