RECAP: Transparent Inference-Time Emotion Alignment for Medical Dialogue Systems

2025-09-12

AI Summary

Large medical foundation models often generate clinically accurate yet emotionally sterile responses, failing to meet patients’ empathetic needs during vulnerable states. To address this, we propose RECAP—a novel inference-time framework that achieves zero-shot emotional alignment via a modular, interpretable five-stage chain: Reflect-Extract-Calibrate-Align-Produce. RECAP integrates appraisal theory with Likert-scale signals to construct a transparent, auditable affective regulation mechanism, requiring no model fine-tuning and supporting multi-scale architectures. Empirical evaluation demonstrates substantial improvements in emotional alignment: +22–28% on emotion benchmarks for 8B-parameter models, and +10–13% for larger models. Clinical expert evaluations further confirm that RECAP-generated responses exhibit significantly enhanced empathic expression compared to strong baselines. This work establishes the first zero-shot, inference-only approach for emotion-aware medical dialogue, offering both interpretability and practical deployability across model scales.

Technology Category

• Cognitive Modeling & Cognitive Systems: Affective Computing
• Natural Language Processing: (Large) Language Models
• Humans and AI: Emotional Intelligence

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large language models in healthcare often miss critical emotional cues, delivering medically sound but emotionally flat advice. This is especially problematic in clinical contexts where patients are distressed and vulnerable, and require empathic communication to support safety, adherence, and trust. We present RECAP (Reflect-Extract-Calibrate-Align-Produce), an inference-time framework that adds structured emotional reasoning without retraining. By decomposing empathy into transparent appraisal-theoretic stages and exposing per-dimension Likert signals, RECAP produces nuanced, auditable responses. Across EmoBench, SECEU, and EQ-Bench, RECAP improves emotional reasoning by 22-28% on 8B models and 10-13% on larger models over zero-shot baselines. Clinician evaluations further confirm superior empathetic communication. RECAP shows that modular, theory-grounded prompting can systematically enhance emotional intelligence in medical AI while preserving the accountability required for deployment.

Problem

• Addresses emotionally flat advice in medical AI
• Enhances empathy without model retraining
• Improves emotional reasoning in clinical dialogues

Innovation

• Uses transparent appraisal-theoretic stages
• Adds structured emotional reasoning without retraining
• Employs modular theory-grounded prompting approach