This paper introduces **CogRAG**, a new RAG pipeline, for conversational mental health support. Different from conventional RAG that is usually optimized for semantic relevance, CogRAG integrates cognitive factors into both retrieval and response generation stages. To do that, CogRAG uses projections by Appraisal-Theory- and CBT-inspired prompts that are applied symmetrically to patient queries and therapist responses, bridging their lexical asymmetry for retrieval. To deal with crisis cases, it augments safety-planning cards from data that guide generation to follow therapies. Experimental results on two datasets show that cognitive modeling helps improve retrieval and generation performance; gains are most pronounced in counseling quality, while safety-related improvements are modest but consistent. We do not claim to implement clinical CBT or to recover ground-truth cognitive states (Code and data: https://github.com/Cog-AI-Lab/CogRAG).

This paper introduces a new multi-agent framework, CMTD (Cognitive Modeling with Traits and Distortions), for multimodal emotion recognition in conversations (MERC). Instead of relying on shallow analysis of emotions, CMTD reconstructs a cognitive model by taking advantage of stable personality traits, dynamic cognitive distortions, visual and acoustic features of interlocutors to enhance the emotional intelligence of LLMs. CMTD includes trait, distortion detection, vision, and speech agents that provide psychological and multimodal indicators for the fusion agent to make the final prediction. Experimental results on MELD and IEMOCAP show that traits temper negativity bias from distortions, and cognitive modeling with psychological, visual, and acoustic information can improve the performance of MERC.CMTD is flexible and easy to adapt to advanced emotional AI systems (Github link: https://github.com/Shaun-le/CMTD.git).

This paper introduces a framework for evaluating structural fidelity in emotion reasoning: the degree to which the causal structure of AI-generated emotion explanations aligns with human-curated explicit causal abstractions. Unlike traditional surface-level text overlap metrics, we propose two metrics: Structural Causal Similarity (SCS) for measuring static structural alignment, and Dynamic Causal Sensitivity (DCS) as a proxy for measuring dynamic sensitivity to counterfactual interventions. We evaluate these on EmoReason-4k, a new benchmark comprising two human-annotated datasets ( each) from social media and conversational dialogues. Testing seven LLMs identifies consistent graph-generation behaviors across models and domains, such as “structural elaboration” (a tendency for models to explicate intermediate steps that humans leave implicit) and sensitivity to shifts in causal agency locus. However, models struggle to generate updated causal states in complex discourse. By evaluating structural fidelity against human-curated causal abstractions, our framework offers a structured method to diagnose the structural compliance of explicit explanations against formalized human abstractions. Ablation studies support that our metrics capture underlying topological properties rather than just parsing artifacts. Data and prompts are available to support reproducibility.

We demonstrate Metamo, a browser-based dialogue system that transforms an off-the-shelf large language model into an empathetic coach for everyday workplace concerns. Metamo introduces a light, single-pass wrapper that first identifies the cognitive distortion behind an emotion, then recognizes the user’s emotion, and finally produces a question-centered reply that invites reflection, all within one model call. The wrapper keeps the response time below two seconds in the API, yet enriches the feedback with cognitively grounded insight. A front-end web interface renders the detected emotion as an animated avatar and shows distortion badges in real time, whereas a safety layer blocks medical advice and redirects crisis language to human hotlines. Empirical tests on public corpora confirmed that the proposed design improved emotion‐recognition quality and response diversity without sacrificing latency. A small user study with company staff reported higher perceived empathy and usability than a latency‐matched baseline. Metamo is model-agnostic, illustrating a practical path toward cognition-aware coaching tools.