Motivation. AI Agents empowered by Large Language Models (LLMs) have shown advanced automation capabilities in executing complex tasks. With the rapid advancement in LLMs' reasoning abilities, AI agents have proven successful in diverse applications such as coding. Recently, Multimodal Foundation Models (MFMs) have emerged as a cutting-edge frontier in artificial intelligence, demonstrating remarkable potential to enhance cognitive and perceptual capabilities in AI Agents. Recent breakthroughs reveal that models integrating visual, textual, and auditory inputs enable fine-grained cross-modal reasoning. These systems excel at parsing complex scenarios, generating context-aware descriptions, and tackling tasks requiring synergistic perception and language understanding—capabilities critical to robotics, human-computer interaction, and beyond. Here, we propose empowering AI agents to operate in multimodal scenarios by leveraging MFMs' advanced multimodal reasoning capabilities.

Background & Application. Recent AI agents have revealed their potential in solving more complex problems under multimodal settings. For example, OS-Copilot is ready to interact with computer OS like a human, such as web browsing, coding, and using third-party applications. Another trending AI agent application is the AI Scientist Agent for research automation, such as designing and running experiments. Nevertheless, existing AI Scientist Agents typically rely on textual input, and the informative multimodal signals are often overlooked. Lastly, Embodied Agents represent a trending area in both research and application, where AI agents interact with the physical (or simulated) world using sensors and robotic arms. Driven by potential applications and deployment scenarios, we define four categories of multimodal AI agents: Digital Agents, Virtual Agents, Wearable Agents, and Physical Agents. Digital Agents are software-based agents that operate exclusively with digital environments where Virtual Agents exist within virtual or simulated environments, such as VR platforms. Wearable Agents are integrated into wearable devices to provide real-time assistance and Physical Agents primarily interact with the physical world. Each category leverages multimodal reasoning—the integration and interpretation of multiple input modalities to perceive, understand, and respond effectively within their respective environments. Challenges. Multimodal reasoning faces unique theoretical and technical challenges compared to unimodal approaches. First, integrating heterogeneous data (text, images, video) demands innovations in model architectures, training paradigms, and evaluation frameworks. The exponential growth in computational complexity due to multimodal data raises concerns about resource efficiency and real-time responsiveness. Second, applying these models to frontier domains—such as multi-agent collaboration, scientific discovery, and embodied intelligence systems—requires overcoming bottlenecks in cross-modal semantic understanding and knowledge transfer. Crucially, the interpretability and robustness of multimodal reasoning systems remain unresolved foundational issues, directly impacting the deployment of reliable real-world applications (e.g., scientific AI agents, bio-inspired robotics). Addressing these challenges necessitates breakthroughs in cross-modal representation alignment, dynamic attention mechanisms, and uncertainty modeling of multi-source information.

Challenges. Multimodal reasoning faces unique theoretical and technical challenges compared to unimodal approaches. First, integrating heterogeneous data (text, images, video) demands innovations in model architectures, training paradigms, and evaluation frameworks. The exponential growth in computational complexity due to multimodal data raises concerns about resource efficiency and real-time responsiveness.

Second, applying these models to frontier domains—such as multi-agent collaboration, scientific discovery, and embodied intelligence systems—requires overcoming bottlenecks in cross-modal semantic understanding and knowledge transfer. Crucially, the interpretability and robustness of multimodal reasoning systems remain unresolved foundational issues, directly impacting the deployment of reliable real-world applications (e.g., scientific AI agents, bio-inspired robotics). Addressing these challenges necessitates breakthroughs in cross-modal representation alignment, dynamic attention mechanisms, and uncertainty modeling of multi-source information.

About this Workshop. All discussions under the scope of multimodal reasoning are welcome. This workshop aims to bring. together researchers with various backgrounds to study the next generation of multimodal reasoning systems. To foster an inclusive dialogue and debate space, we invite speakers and panelists from diverse backgrounds and areas of expertise. Our roster includes both renowned researchers and emerging investigators who have driven promising advances in the field.