Top 10 Open Challenges Steering the Future of Diffusion Language Models and Its Variants, presented by Huawei

The paradigm of Large Language Models (LLMs) is currently defined by auto-regressive (AR) architectures, which generate text through a sequential “brick-by-brick” process. Despite their success, AR models are inherently constrained by a causal bottleneck that limits global structural foresight and iterative refinement. Diffusion Language Models (DLMs) offer a transformative alternative, conceptualizing text generation as a holistic, bidirectional denoising process akin to a sculptor refining a masterpiece. However, the potential of DLMs remains largely untapped as they are frequently confined within AR-legacy infrastructures and optimization frameworks. In this Perspective, we identify ten fundamental challenges—ranging from architectural inertia and gradient sparsity to the limitations of linear reasoning—that prevent DLMs from reaching their “GPT-4 moment.” We propose a strategic roadmap organized into four pillars: foundational infrastructure, algorithmic optimization, cognitive reasoning, and unified multimodal intelligence. By shifting toward a “diffusion-native” ecosystem characterized by pyramid tokenization, active remasking, and latent thinking, we can move beyond the constraints of the causal horizon. We argue that this transition is essential for developing next-generation AI capable of complex structural reasoning, dynamic self-correction, and seamless multimodal integration.

Hierarchical Reasoning Model: A Brain-Inspired Path to AGI, presented by Sapient

This presentation introduces the Hierarchical Reasoning Model (HRM), a next-generation, brain-inspired hierarchical architecture developed by Sapient Intelligence

Physical AI Ecosystem: Tackling the Key Barriers to Embodied Intelligence, presented by Alibaba

Embodied intelligence has long been seen as a necessary step on the path toward Artificial General Intelligence (AGI). Looking back at recent progress, we can clearly see that robotics is on track to become the next foundational infrastructure after the internet and large-scale foundation models. Yet despite this promise, truly autonomous robots and embodied agents remain rare in production today.

In this talk, we argue that the key blockers are no longer just algorithms, but a set of tightly coupled barriers across systems, data, models, and talent. We will use the RynnBot embodied intelligence platform, the RynnRCP protocol, the Rynn family of open-source embodied foundation models, and our collaboration cases with industry and academia to illustrate how we approach the four barriers in a coordinated way. By co-creating a Physical AI ecosystem, we aim to work with upstream and downstream partners and industry stakeholders to collectively usher in “the decade of robotics.”

Xiaomi x AAAI 2026: Empowering Tomorrow’s AI Visionaries, presented by Xiaomi

Bringing together an elite roster of tech leaders from Xiaomi’s core AI teams, this exclusive workshop will feature Xiaomi’s latest technological breakthroughs and elaborate on the framework of its premium talent recruitment initiative and bespoke offerings.

Artificial Intelligence at Allianz Technology, presented by Allianz

Allianz is a global insurance and asset management group operating across diverse markets and regulatory environments. This talk introduces how Allianz Technology approaches the development and use of AI systems in such a large, complex enterprise, focusing on the technical context, the role of AI across the software development lifecycle, and perspective on how this work connects to business value.

Accelerating AI Research with AWS Trainium: High-Performance Deep Learning Training at Scale, presented by Amazon

Join us for an intensive hands-on workshop designed for AI researchers exploring cutting-edge deep learning training on AWS Trainium—Amazon’s purpose-built AI accelerator chip. This workshop will equip you with practical skills to leverage Trainium’s powerful capabilities for training state-of-the-art AI models at scale.

**Attendees should bring a Wi-Fi enabled laptop

What You’ll Learn:

• AWS Trainium Architecture: Understand the unique systolic array design optimized for advancing state-of-the-art AI ideas and applications
• AWS Neuron SDK: Master the tools and frameworks for distributed training and inference on Trainium
• Performance Optimization: Learn best practices for maximizing training efficiency and scaling across large distributed systems
• Build on Trainium Grant Opportunity: Participants will learn about the $110 million Build on Trainium program—AWS’s investment in AI research and university education.

AI x Amex: Application and Research, presented by American Express

A talk around AI and Gen AI at Amex on industry application and theoretical research where Amex will be sharing an overview of their business, as well as some of the data science use cases.

Engineering Agentic Intelligence: A Pipeline from Efficient Reasoning to Multimodal Grounding to Agentic Collaboration, presented by The Chen Institute

The rapid evolution of large language models from text generators into agentic systems that plan, reason, and interact with tools marks a fundamental shift in AI capabilities. However, turning this research potential into robust, deployable intelligence confronts three critical engineering bottlenecks: the exploding computational cost of long reasoning traces, the instability of training agents for multi-step tool use, and the challenge of grounding agent decisions in rich, multimodal worlds like images and long videos. Ourtutorial addresses these interconnected challenges head-on, presenting a unified engineering framework that moves from diagnosing and optimizing the core reasoning engine to embodying it in complex visual and temporal environments. We synthesize five cutting-edge research threads into a coherent blueprint, teaching attendees how to build efficient, stable, and grounded agents—bridging the gap between groundbreaking agentic research and practical, scalable deployment.