The early and accurate diagnosis of Alzheimer's Disease (AD) is essential for effective intervention, yet remains challenging due to the complexity of clinical data. While large language models (LLMs) have become a promising avenue for medical diagnostics, existing approaches often overlook the implicit medical knowledge embedded in diagnostic labels, limiting diagnostic reliability and interpretability. Motivated by this, we introduce Label-Aware Multi-Agent Alzheimer's Disease Diagnosis with Counterfactual Reasoning (LAMA-AD), a labelaware multi-agent framework that explicitly incorporates clinical priors into the diagnostic reasoning process and integrates both factual and counterfactual reasoning. Specifically, LAMA-AD consists of two components: (1) Label-Aware Multi-Agent Diagnosis, which deploys multiple independent analyst agents, each guided by distinct label-aware prior knowledge, to explore diverse diagnostic hypotheses in parallel. Each agent provides supporting facts for its reasoning process, thereby enhancing both the depth and robustness of the analysis. (2) Counterfactual Reasoningsystematically assesses conflicting hypotheses to yield stable, interpretable, and highly accurate diagnostic conclusions. Experimental verification confirms LAMA-AD's superiority over established methods. Analysis further shows it generates semantically rich, logically consistent explanations.

Accurate diagnosis of Alzheimer’s disease is vital for effective treatment. Recently, a mainstream approach emerged that uses large language models (LLMs) to generate interpretable diagnosis. However, these methods face two significant challenges: (1) Incomplete Utilization of Raw Image Information: Converting images into text for LLMs input leads to the loss of the original image information. (2) Hallucination in LLMs: Diagnostic biases from LLMs hallucinations are neglected. To address these challenges, we propose a Medical Diagnosis Chain-of-Thought with Self-Diagnostic Refinement (MdCoT) framework. The MdCoT framework consists of two core modules: (1) Multimodal AD Diagnostic Chain-of-Thought: By including raw images as input, multimodal large language models (MLLMs) are allowed to capture fine-grained features, and (2) Self-Diagnostic Refinement: By guiding MLLMs to self-check and refine the diagnostic results, MdCoT achieves to mitigate hallucination. Experiments on widely used benchmarks demonstrate that MdCoT surpasses all baselines. Besides, extensive auxiliary experiments demonstrate the superiority of MdCoT.

Recently, Profile-based Spoken Language Understanding (SLU) has gained increasing attention, which aims to incorporate various types of supplementary profile information (i.e., Knowledge Graph, User Profile, Context Awareness) to eliminate the prevalent ambiguities in user utterances. However, existing approaches can only separately model different profile information, without considering their interrelationships or excluding irrelevant and conflicting information within them. To address the above issues, we introduce a Heterogeneous Graph Attention Network to perform reasoning across multiple Profile information, called Pro-HAN. Specifically, we design three types of edges, denoted as intra-Pro, inter-Pro, and utterance-Pro, to capture interrelationships among multiple Pros. We establish a new state-of-the-art on the ProSLU dataset, with an improvement of approximately 8% across all three metrics. Further analysis experiments also confirm the effectiveness of our method in modeling multi-source profile information.