Industry Talks

Abstract

Large models are moving from "can they get smarter" to "can stronger intelligence enter the real world at lower cost and higher reliability." This talk proposes the Intelligence-Efficiency Ratio: the real, reliable, professional, and sustainable intelligence value delivered per unit of total resources invested. Drawing on practice with the Ant Ling model families, it shows how three mutually reinforcing paths - efficient architectures, real-task evaluation, and an open ecosystem - form a flywheel that brings high-level intelligence to more people, organizations, and industries at scale.

Biography

Dr. Jun Zhou is Vice President of Ant Group. His research focuses on foundation models, machine learning, and other AI technologies, as well as their industrial applications. He has published over 100 papers at top-tier AI conferences, with an H-index of 60, and has received three international best paper awards and 46 granted U.S. patents.

As an early member of Alibaba Cloud's technical team, he contributed to the development and large-scale commercial deployment of Apsara, Alibaba's independently developed cloud computing operating system. As one of the founding members of AI at Alibaba and Ant Group, he has been deeply involved in and led major projects including personalized recommendation for the Double 11 shopping festival, the Platform for AI (PAI), the Ant Graph Learning system (AGL), AI-based risk management, and the intelligent evolution of Alipay. Under his leadership, his teams have ranked first in four international algorithm competitions. His honors include the First Prize of the Ministry of Education Award for Outstanding Scientific Research Achievements, the CCF Science and Technology Progress Award (First Prize), and the Zhejiang Province Science and Technology Progress Award (First Prize), among others.

Abstract

AI Agents are evolving toward autonomous collaborative systems. This lecture investigates multi-agent collaboration and inter-agent communication mechanisms, with Mindline serving as a human-AI shared thought container. The agent Echous is deployed on the EchousGrid network, enabling inter-agent coordination via structured communication protocols. We further elaborate the prospective paradigm in which agents interface with human cognitive space and the physical world. Open challenges regarding cognitive-physical integration are analyzed, and a forward-looking framework is proposed for an interconnected intelligent ecosystem integrating multi-agent coordination, human-AI shared cognition and physical-world execution.

Biography

Jie Li, formerly with Shaanxi Normal University. General Manager, Zhengzhou Zhongye Technology Co., Ltd., Standing Committee Member, CCF Technical Committee on Natural Language Processing, Director, Institute of Artificial Intelligence, National Supercomputing Zhengzhou Center, Co-Director, Zhengzhou University–Zhongye NLPLab, Committee Member, Henan Technical Committee for Standardization of Artificial Intelligence Medical Devices, Committee Member, Henan Technical Committee for Standardization of Robotics. Principal drafter of White Paper on Artificial Intelligence Standardization (2021), Knowledge Graph Selection and Implementation Guide (2021), Guidelines for the Standardization of Artificial Intelligence Ethics Governance (2023).

Abstract

In this talk, we will share our perspectives on the evolving landscape of AI, tracing its progression from understanding and generation toward interactive intelligence and real-world action. We will introduce our recent advances in JoyAI-VL-Interaction, JoyAI-Video-Edit, and JoyAI-Echo, showcasing our explorations in multimodal intelligence, interactive content creation, and human-AI interaction. We will also highlight JD's efforts in embodied AI data collection, emphasizing the importance of high-quality real-world data in enabling AI systems to learn, interact with, and act in physical environments. Together, these developments reflect our vision for the next generation of AI systems: moving beyond understanding and generating content toward more natural interaction and embodied intelligence.

Biography

Jiaqi Wang heads the Natural Language Processing Lab at JD Explore Academy, leading the development of large language models (LLMs) and vision-language models (VLMs). He has published over 70 papers in top-tier international journals and conferences, with more than 30,000 citations on Google Scholar. Several of these papers have been recognized among the year's most influential papers at top-tier conferences. He has also served as an Area Chair for NeurIPS, ICLR, ICML, and CVPR.

Abstract

In recent years, the technology of AI large models has been advancing rapidly, and it has been widely applied in the fields of text, images, and videos. However, its practical application in the audio domain still has a considerable room for development. Due to the higher difficulty in obtaining high-quality audio data compared to other fields, traditional methods usually involve training multiple models in the audio field and conducting complex system integration to complete specific sound scene tasks. Clearly, such methods are costly and have low reusability in practical applications. Therefore, building sound AI composite models based on the existing AI large models and applying the models to audio scenarios has become an significant direction worthy of research and exploration.

Biography

Dr. Ye Liu currently serves as the Vice President of Technology of Lizhi Inc. under Sound Group Inc., a Nasdaq-listed company. He directs the R&D teams for Artificial Intelligence(AI), Big Data, Business Technology, Fundamental Research, and Audio Systems. With extensive experience in the industrial deployment of artificial intelligence and big data, he possesses deep expertise in applying large-scale AI models to audio-domain internet scenarios. He has authored over 20 papers in international conferences and journals, holds more than 20 granted patents, and demonstrates profound proficiency in sound AI technology research and industry scenario applications.

Abstract

As large language models move toward sustained interaction and long-running tasks, memory is becoming essential infrastructure for personalized, continuous, and adaptive AI agents. This talk presents two complementary efforts at MemTensor: MemOS and MemT 1.0. MemOS unifies plaintext, activation, and parametric memory as system resources with coordinated management, scheduling, and governance, covering their lifecycle from writing and retrieval to updating, migration, and forgetting. MemT 1.0 embeds memory into foundation models through persistent, dynamically evolving native memory states. It maintains and accesses historical memory through forward computation alone, with frozen weights and no backpropagation. The talk will examine selective memory writing, compression and capacity limits, knowledge conflicts, the stability–plasticity trade-off, long-horizon evaluation, training and inference efficiency, and privacy, security, and auditable governance. It will conclude with an outlook on the co-evolution of memory operating systems and memory-native models.

Biography

Zhiyu Li, Ph.D., is Co-founder and CTO of MemTensor and Technical Lead and Researcher at the Large Model Center of the Institute for Advanced Algorithms Research, Shanghai. He is a Shanghai Oriental Talent Program (Top Talent) honoree and a 36Kr “36Under36” Science and Technology Innovator. His research focuses on LLM pretraining, memory augmentation, efficient evaluation, and applications. He previously led core algorithm teams at Alibaba and Xiaohongshu. With his team, he introduced Memory³, a model with hierarchical memory, in 2024; MemOS, a memory operating system for LLMs, in 2025; and MemT 1.0, a memory-native foundation model, in 2026. MemOS has attracted over 11,000 GitHub stars and more than 20,000 developers. Related technologies have been deployed at Bank of China, China Merchants Securities, China Telecom, and Xinhua News Agency. Li has published over 70 papers in venues including Patterns, NeurIPS, ICLR, ACL, and TKDE, and holds more than 10 granted patents. He also authored the Chinese-language monograph Memory Engineering: Building Agent Memory Systems in Practice. His research has been featured in a cover story in MIT Technology Review.

Abstract

Imagine a personal agent that does more than recap your day. It remembers your commitments, tracks unresolved dependencies, and notices that a delayed confirmation could put next week’s delivery at risk. Without waiting for you to ask, it flags the risk. It then joins the project group chat, with your permission, to check on the dependency, help the team agree on next steps, and follow up until the issue is resolved.

We believe memory, proactive assistance, and lifelong learning are essential to building such agents. Beyond personal assistance, we’ll explore how agents learn the context of ongoing group discussions and use that understanding to support collaboration. Join us as we explore what it takes to build agents that remember, anticipate, and collaborate.

Biography

Zenan Xu, Senior Researcher in the Foundation Model Department at Tencent Hunyuan, works on large language models and personal agents. His current work focuses on building agents that understand users over time and provide proactive support, with an emphasis on memory and lifelong learning. He joined Tencent in 2023, and his work has included LLM post-training, using synthetic data generation and reinforcement learning to improve instruction following, long-context understanding, and complex multi-turn interactions. He received his Ph.D. in Computer Science from Sun Yat-sen University in 2023. Previously, he was a research intern at Microsoft Research Asia and Huawei Noah’s Ark Lab, where he worked on natural language processing and pre-trained models.