RETHiNK: Simulating Human Reasoning via Multi-LLM Debate for Cognitive Reframing of Negative Thoughts

Cognitive Reframing, a core technique in Cognitive Behavioral Therapy (CBT), involves transforming negative thoughts into more balanced perspectives. While traditionally facilitated through therapist–client interaction, growing research interest has turned toward modeling this process computationally to generate quality reference reframes. Previous studies using Large Language Models (LLMs) for cognitive reframing often overlook the structured reasoning processes emphasized in CBT. To address this, we propose RETHiNK, a psychologically informed NLP framework that emulates structured reasoning through a multi-LLM debate pipeline. It operates in three phases: Preprocessing, which identifies emotions, intentions, and cognitive distortions to contextualize user input; Debate, which evaluates thoughts, infers evidence, and proposes actions via multi-LLM debate grounded in psychological principles; and Refinement, which iteratively enhances reframes with LLM-based feedback, introducing an overpositivity check for quality control. Experimental results demonstrate that RETHiNK outperforms the baselines in human evaluation across helpfulness, empathy, and rationality. This work is conducted entirely in an offline research setting with non-clinical data and is not intended for therapeutic or clinical use. Our goal is to investigate how structured reasoning can improve reframing generation, rather than building a deployable mental health system.