Large language models exhibit significant potential for psychological support, yet they often generate fragmented and emotionally inconsistent dialogues that lack the therapeutic structure necessary for reliable assessment.To address these issues, we introduce **VeilEval**, a clinically grounded and privacy-preserving benchmark equipped with interpretable metrics for evaluating multi-turn psychological dialogues.Furthermore, we propose Emotion-Resonance (**EmoRes**), a multi-agent framework that boosts psychological reasoning via a Topic-Mining Emotional Agent and a multi-perspective Self-Reflection Agent, thereby jointly improving topic continuity, emotional coherence, and clinical interpretability.Experiments demonstrate that EmoRes achieves up to ∼ 3× improvement over strong baselines on VeilEval, with its effectiveness further validated by ablation studies and human evaluations.

Multimodal large language models (MLLMs) have demonstrated strong performance across diverse multimodal tasks, achieving promising outcomes. However, their application to emotion recognition in natural images remains underexplored. MLLMs struggle to handle ambiguous emotional expressions and implicit affective cues, whose capability is crucial for affective understanding but largely overlooked. To address these challenges, we propose MERMAID, a novel multi-agent framework that integrates a multi-perspective self-reflection module, an emotion-guided visual augmentation module, and a cross-modal verification module. These components enable agents to interact across modalities and reinforce subtle emotional semantics, thereby enhancing emotion recognition and supporting autonomous performance. Extensive experiments show that MERMAID outperforms existing methods, achieving absolute accuracy gains of 8.70%–27.90% across diverse benchmarks and exhibiting greater robustness in emotionally diverse scenarios.