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

1. Problem Statement

Large language models (LLMs) in healthcare often generate responses that are emotionally flat, opaque, or fail to provide the transparent reasoning required for clinical trust. While current AI-assisted medical dialogue systems excel at knowledge retrieval, they treat emotional reasoning as a "black box." This opacity is critical in healthcare because clinicians cannot deploy systems where the rationale for an emotional response is hidden, nor can they determine if the model recognizes a patient's psychological state or merely pattern-matches to empathetic-sounding language. Furthermore, existing training-time methods (e.g., RLHF, supervised fine-tuning) embed empathetic behavior into model weights, making the reasoning process uninterpretable and difficult to audit.

2. Methodology: The RECAP Framework

The authors propose RECAP (Reflect–Extract–Calibrate–Align–Produce), an inference-time framework grounded in cognitive appraisal theory. Unlike training-time approaches, RECAP requires no retraining; it operates by decomposing patient input into auditable, appraisal-theoretic stages during a single conversation turn. The pipeline externalizes emotional reasoning into five explicit stages, producing inspectable intermediate artifacts:

1. Reflect (Situation Abstraction): The model distills raw patient input into a concise description of the core situation, stripping lengthy narratives to capture essential circumstances and emotional triggers.
2. Extract (Latent Factor Induction): The model identifies three psychological factors shaping the emotional response (e.g., Control, Support, Uncertainty). These map to primary appraisal (goal relevance) and secondary appraisal (agency, coping potential) dimensions. The model assigns values (e.g., Low, Moderate, High) and justifications to these factors.
3. Calibrate (Candidate Emotion Extraction & Likert Rating):
   • Extraction: The model generates 5–7 plausible candidate emotions rather than a single label, acknowledging mixed affect.
   • Calibration: Each candidate is rated on a five-point Likert scale (mapped to probabilities 0.05–0.95). This step forces the model to express uncertainty and partial belief rather than guessing, creating a clinician-reviewable vector of emotion likelihoods.
4. Align (Response Generation): Guided by the emotion-likelihood profile, the model generates a response that acknowledges the highest-probability feelings, validates them, and adjusts tone before offering medical guidance. This follows clinical frameworks prioritizing emotional validation before information delivery.
5. Produce: The final output is the emotionally aligned response, conditioned on the structured intermediates.

The framework is model-agnostic and designed to produce "auditable checkpoints" (situation abstractions, factor ratings, emotion profiles) that clinicians can inspect, verify, or override.

3. Key Contributions

• Inference-Time Decomposition: RECAP introduces a structured, five-stage pipeline that externalizes emotional reasoning without fine-tuning, allowing for the inspection of how models reason about affect.
• Appraisal-Theoretic Grounding: By grounding the pipeline in cognitive appraisal theory, the method moves beyond surface-level stylistic mimicry to explicit modeling of goal relevance, agency, and coping potential.
• Quantifiable Intermediates: The framework generates discrete artifacts (factor values, Likert ratings) that enable statistical analysis of model behavior, revealing systematic biases (e.g., underweighting relational factors like social support).
• Modular Design: The pipeline supports clinical override, as emotional guidance can be adjusted based on external input before the final response is generated.

4. Experimental Results

The authors evaluated RECAP across multiple benchmarks (EmoBench, SECEU, EQ-Bench3, HealthBench) and models ranging from 8B to 120B parameters, comparing it against Zero-Shot, Emotion Priming, and Supportive Clinician Persona baselines.

• Emotional Intelligence Benchmarks:
  • RECAP consistently improved alignment with human judgments. On the SECEU benchmark, it reduced the distance from human consensus by 20% and achieved the highest EQ scores across all model scales.
  • On EmoBench, RECAP showed a 28% relative improvement for the 8B model compared to Zero-Shot, while larger models (120B) saw smaller gains (approx. 5%), suggesting structured reasoning compensates for capabilities that larger models have internalized.
  • On EQ-Bench3, RECAP achieved significant Elo gains (e.g., +70 points for the 8B model).
• Medical Communication Quality (HealthBench):
  • RECAP improved overall quality scores, with the largest gains in Completeness and Communication Quality.
  • A trade-off was observed: Accuracy and Instruction Following scores decreased slightly, reflecting a prioritization of comprehensive, emotionally tuned responses over literal precision.
• Clinical Evaluation (Blinded Human Study):
  • Two board-certified oncology fellows rated RECAP responses significantly higher than baselines (3.41 vs. 2.93–2.95, p < .001).
  • RECAP achieved 76–88% win rates against baselines in single-turn scenarios.
  • Qualitative analysis showed RECAP responses were characterized as "emotionally cognizant, supportive and complete," whereas baselines were often criticized as "robotic."
• Systematic Biases: Intermediate outputs revealed that models systematically underweight relational factors such as social support, a bias detectable only through RECAP's structured intermediates.

5. Significance and Claims

The paper claims that RECAP demonstrates that principled prompting can enhance medical AI's emotional intelligence while maintaining the transparency required for clinical deployment.

• Transparency as a Prerequisite: The authors argue that in healthcare, explainability is not merely desirable but essential for safety, regulatory compliance, and clinical trust. RECAP addresses this by exposing intermediate artifacts that clinicians can inspect.
• Scalability of Reasoning: The results suggest that explicit decomposition is particularly valuable for smaller models, enabling them to achieve performance comparable to much larger models without retraining.
• Limitations: The authors modestly note that RECAP introduces computational overhead (3–5× token usage), is text-only (missing tone/pauses), and is currently a research contribution rather than a deployable clinical tool. They caution against deployment without extensive safety testing and regulatory review, particularly in high-acuity scenarios where directive action is prioritized over empathy.

In conclusion, RECAP transforms opaque pattern-matching into a transparent pipeline, providing a path toward AI systems that are both more empathetic and more auditable.