CARE: Mitigating Knowledge Intrusion for Empathetic Dialogue via Intent-Gated Retrieval and Conflict-Aware Reasoning

ACL ARR 2026 January Submission3418 Authors

Keywords: Empathetic Dialogue, Retrieval-Augmented Generation, Large Language Models

Abstract: Retrieval-Augmented Generation (RAG) can undermine empathetic dialogue when retrieved content is contextually or emotionally misaligned, leading the model to uncritically rely on retrieved documents as its response—a failure mode we term “Knowledge Intrusion.” To mitigate this, we propose CARE (Conflict-Aware Reasoning for Empathy), which synergizes Intent-Gated Retrieval and Latent Critique to ensure relevance, reinforced by Conflict-Aware DPO to enhance robustness against noisy contexts. Experiments on EmpatheticDialogues and ESConv demonstrate that CARE outperforms strong baselines, achieving F1 score gains of 7.3\%--33.1\% while maintaining high robustness, evidenced by a Context Rejection Score (CRS) exceeding 70\%. Our code is available at https://anonymous.4open.science/r/CARE-FA8B.

Paper Type: Long

Research Area: Human-AI Interaction/Cooperation and Human-Centric NLP

Research Area Keywords: human-AI interaction/cooperation, human-centered evaluation, conversational modeling

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 3418