OCTO-CARES: Towards Human-like Emotional Support Conversation with Large Language Models

As large language models (LLMs) become increasingly integrated into everyday life, they are also being used for daily emotional support. While LLMs provide reassurance and guidance through empathetic language and structured support strategies, some users may perceive these responses as formulaic and impersonal. These users may seek the sense of connection that comes from shared understanding and perspectives shaped by lived experience and interpersonal relationships, qualities that current LLMs can struggle to convey. Therefore, in this work, we aim to improve the naturalness and interpersonal resonance of LLM-generated emotional support by matching users’ support needs with relevant human peer-support exemplars. Specifically, we propose Octo-Classifier Attention-Weighted Retrieval for Empathetic Support (Octo-CARES), which combines a classifier that identifies users' support needs across eight help-seeking patterns with an attention-weighted retrieval mechanism that selects relevant human peer-support exchanges. We construct Octo-Pairs, a dataset of 2,653 naturally occurring Reddit post–comment pairs annotated with these patterns. Each pair links a concrete help-seeking situation to a highly upvoted peer response. The retrieved exchanges are incorporated into the LLM's context as exemplars of understanding and experiential insight to guide response generation. Experiments on both in-domain and out-of-domain datasets demonstrate that Octo-CARES helps LLMs generate more supportive and human-like responses. Our code is currently available via an anonymous link./r/OCTO-CARES-Towards-Human-like-Emotional-Support-Conversation-with-Large-Language-Models-417C/