The rapid advancement of AI is fundamentally transforming how scientific research is conducted. AI Co-Scientist systems—intelligent multi-agent frameworks designed to collaborate with human scientists—represent a paradigm shift from AI as a tool to AI as a research partner. These systems can analyze vast literature corpora, generate novel hypotheses, design and execute machine learning experiments, and provide continuous feedback, all while maintaining the "scientist-in-the-loop" paradigm essential for scientific rigor. This workshop focuses specifically on AI Co-Scientists for Machine Learning and AI research, rather than the broader domain of AI for natural sciences. We are interested in systems that help advance ML/AI science itself: developing better models, algorithms, evaluation methodologies, and empirical understanding through tightly coupled human–AI collaboration.

The workshop will bring together Amazon scientists and external researchers from academia and industry to:

• Share emerging architectures, design patterns, and best practices for agentic AI Co-Scientist systems
• Discuss challenges such as reliability, credit assignment, evaluation, and reproducibility in AI-assisted research
• Explore case studies where AI agents contribute meaningfully to hypothesis generation, experimentation, and iterative discovery
• Identify opportunities for collaboration and open research directions in this rapidly evolving space

Speakers: Peter Clark, Senior Research Director, Allen AI; James Zou, Associate Professor, Stanford; Chenhao Tan, Associate Professor, University of Chicago; Wei Wang, Professor, UCLA

Organizing committee: Olcay Boz, Principal Scientist, Amazon; Amila Weerasinghe, Sr. Applied Scientist, Amazon; Utkrisht Rajkumar, Sr. Applied Scientist, Amazon

Multimodal LLMs have revolutionized tasks like visual question answering, image captioning, video understanding, and cross-modal reasoning. However, advancement in multimodal AI capabilities typically comes at the expense of increasing computational and memory demands, creating significant deployment challenges, especially for edge computing and privacy-sensitive applications. As we move towards ambient intelligence and edge-based AI processing, the need for efficient architectures that enable fast, cost-effective inference under given memory and computation constraints becomes increasingly critical.

This workshop provides a comprehensive summary of emerging strategies for building efficient multimodal systems that maintain high performance while substantially reducing computational overhead. Through technical presentations and posters, we will explore novel architectural frameworks for efficient multimodal LLMs along with training and optimization techniques for edge deployment. We will also discuss strategies for model compression (pruning, quantization) and scalable approaches to synthetic data collection and benchmarking.

By examining recent advances in efficient multimodal AI, this workshop will highlight how architectural design choices, inference optimization techniques, and training methodologies influence performance across diverse tasks and deployment settings.

Speakers: To be announced soon.

Organizing committee: Ipshita Bhattacharya, Sr. Applied Scientist, Amazon; Daniel Griffin, Sr. Applied Scientist, Amazon; Sankalp Dayal, Applied Science Manager, Amazon; Ananth Ranganathan, Principal Applied Scientist, Amazon

Software development is entering a new phase driven by collaborative AI agents that go far beyond traditional code completion, engaging in planning, implementation, testing, debugging, and documentation alongside human developers and other agents. Recent advances in LLMs, combined with the rapid deployment of AI-powered development tools in production environments, have created an inflection point where design choices about human–AI and agent–agent collaboration will shape the next decade of programming practice.

This workshop brings a dedicated focus on the collaborative dimensions of AI-driven software development, addressing how agent architectures should be designed, how effective human–AI interaction patterns can be established while preserving human agency, and how trust, reliability, and verification can be ensured in real-world deployment. By moving beyond isolated model performance and tackling collaboration as a first-class research problem, the workshop fills a critical gap in current research and provides a timely forum to establish foundational frameworks, system designs, and evaluation methodologies for next-generation, human-centric code development systems.

Speakers: Graham Neubig, Associate Professor, CMU; Eddie Aftandilian, Principal Researcher, GitHub Next; Varun Kumar, Senior Applied Science Manager, Kiro Science, Amazon

Organizing committee: Behrooz Omidvar-Tehrani, Sr. Applied Scientist, Amazon; Luke Huan, Sr. Principal Scientist, Amazon; Shweta Garg, Sr. Applied Scientist, Amazon; Narayanan Sadagopan, Principal Applied Scientist, Amazon

The intersection of AI and security has become increasingly important from two perspectives. On the one hand, the rapid emergence of agentic AI systems has introduced new security challenges that demand urgent attention. Ensuring these powerful systems are safe, robust, and trustworthy is now a top priority. On the other hand, AI is also driving significant progress in the security domain, enabling more advanced threat detection, defense mechanisms, and protection capabilities.

This workshop brings together external academic researchers and Amazon’s internal research communities for talks, poster sessions, and collaborative discussions that explore both dimensions of AI and security. It offers a unique opportunity for Amazon to engage with the broader academic ecosystem to address pressing and emerging challenges in this rapidly evolving field. At the same time, the workshop provides a platform for researchers to exchange new ideas, share state-of-the-art work, and explore meaningful collaboration opportunities.

Speakers: Dawn Song, Professor, UC Berkeley; Vitaly Shmatikov, Professor, Cornell; Baris Coskun, Senior Principal Scientist, Amazon

Organizing committee: Ding Wei, Applied Science Manager, Amazon; Angela Chow, Sr. Security Engineer, Amazon

Speakers: Ion Stoica, Professor, UC Berkley; Minjia Zhang, Assistant Professor, UIUC; Sophia Shao, Associate Professor, UC Berkeley; Ron Diamant, VP and Distinguished Engineer, Trainium, Amazon

Organizing committee: Yida Wang, Principal Applied Scientist, Amazon; Youngsuk Park, Sr. Applied Scientist, Amazon