Developing effective and reliable health prediction models traditionally requires close collaboration between clinicians and machine learning experts, which is costly and limits efficiency. Recent advances in large language model (LLM) agents offer the potential for clinicians to autonomously develop prediction models through natural language interaction. However, existing agent-based systems remain limited in handling heterogeneous health data, complex prediction tasks, and the reliability requirements of healthcare applications. In this work, we present AutoHealth, an uncertainty-aware multi-agent framework for autonomous health prediction. Given task descriptions, AutoHealth dynamically constructs specialized agents that collaboratively perform data exploration, strategy design, code generation, model training, evaluation, and iterative refinement. Beyond optimizing predictive performance, the framework explicitly models uncertainty and generates comprehensive reports to support trustworthy clinical decision-making. To enable robust evaluation, we curate a real-world benchmark comprising 17 tasks across diverse health data modalities and learning settings. On this benchmark, AutoHealth completes all tasks and outperforms state-of-the-art baselines in both prediction performance and uncertainty estimation. These results demonstrate the strong capability of LLM agents in automated health data modeling.

Dominant multimodal emotion recognition paradigms often neglect the intrinsic geometric structure of affect, resulting in representations heavily entangled with non-affective factors. To address this, we propose a Canonical Disentangled Multimodal Generative Framework aimed at recovering the canonical affective manifold from raw data. We explicitly decompose the latent space into a canonical Shared Affective Subspace (z_vad) and a Private Modality Subspace (z_priv). We facilitate this factorization through Supervised Manifold Anchoring and Cross-Modal Manifold Alignment. Experiments demonstrate that our model effectively disentangles affect from private attributes (e.g., identity), achieving superior robustness in zero-shot cross-domain transfer compared to fully supervised baselines, while enabling controllable emotion generation.

This paper describes the system designed by the Baidu PGL Team which achieved the first place in the TextGraphs 2020 Shared Task. The task focuses on generating explanations for elementary science questions. Given a question and its corresponding correct answer, we are asked to select the facts that can explain why the answer is correct for the question and answering (QA) from a large knowledge base. To address this problem, we use a pre-trained language model to recall the top-K relevant explanations for each question. Then, we adopt a re-ranking approach based on a pre-trained language model to rank the candidate explanations. To further improve the rankings, we also develop an architecture consisting both powerful pre-trained transformers and GNNs to tackle the multi-hop inference problem. The official evaluation shows that, our system can outperform the second best system by 1.91 points.