Turning Cutting-Edge AI into a Cocktail at Zhejiang University’s Academic Bar

On the afternoon of May 25, Alibaba International set up an “Academic Bar” at Zhejiang University’s Shao Science Hall, partnering with the FIT forum of the School of Information to host a series of deep‑dive talks focused on emerging AI agent technologies.

Multilingual E‑LLM : AliExpress search algorithm lead Xiao Rong described how his team built a “Multilingual E‑LLM” foundation by pre‑training on more than 500 billion multilingual e‑commerce sentences, reconstructing cross‑language semantic space to solve intent distortion in generative retrieval. He emphasized that global e‑commerce search requires millisecond‑level language bridging to capture true user intent.

Recommendation Paradigm Shift : Ling Feng, head of AliExpress recommendation algorithms, outlined a two‑year evolution: the Sigma large model (SIGIR’26) injects world knowledge into recommendation decisions, enabling understanding of holidays, climate, and cultural differences across 200+ countries; the ASGR generative recall surpasses traditional vector search; the SORT transformer‑based ranking improves order volume by 6 % while doubling inference throughput; and the GPSD generative pre‑training pipeline (KDD’25) further enhances performance.

AI‑Powered Optical Design : Professor Wang Kawei (School of Optoelectronics) presented an “optical design Agent” that uses GPU‑accelerated parallel ray tracing and million‑scale structure search to autonomously reason about lens designs, demonstrating AI’s potential in precision industrial applications.

Safety Foundations : Researcher Hu Jia‑cong (School of Computer Science) highlighted a comprehensive safety‑assessment framework that covers planning, action, and feedback stages of agents, warning that as agents grow stronger, mechanisms are needed to ensure they “tell the truth and act truthfully.”

E‑commerce Agent Benchmark : Chui Yi, head of Lazada AI shopping assistant, introduced the ShoppingBench benchmark containing 2.5 million real product entries and four categories of complex shopping queries. By synthesizing tool‑call trajectories for post‑training, the team improved LLM accuracy on intricate recommendation tasks, and their multi‑agent ProductResearch framework distilled high‑quality long‑range research traces into a single model, yielding notable metric gains.

Virtual User for Market Research : Chen Pei (AI Institute) described constructing high‑fidelity virtual users via persona graphs that retain stable values and memories, enabling low‑cost, scalable user insight and reducing interview expenses from thousands of yuan per hour to a feasible business need.

Agent Harness Architecture : Xie Si‑cong, leader of Accio AI, detailed the Agent Harness design: Skills act as the brain, Subagents as the team, Connectors as the eyes, CLIs as the hands, and Hooks as the heartbeat. This precise goal‑management and context‑orchestration system supports both “stock” and “scheduling” collaboration patterns, achieving 24/7 autonomous multi‑agent operation. He also demonstrated a post‑training pipeline where reinforcement learning continuously extracts execution trajectories to evolve the agentic model’s logical perception.

Edge‑Side Large Models : Doctoral candidate Yan Yuxuan (School of Information) proposed a device‑coordinated joint fine‑tuning approach that mitigates cloud cost and privacy bottlenecks. Experiments on long‑video retrieval show multi‑step reasoning that surpasses cloud models, offering a lighter, more private, and smarter paradigm for the era of ubiquitous intelligent devices.

The event illustrated how academic research on AI agents, multilingual models, recommendation systems, safety, and edge computing is rapidly translating into industrial practice, fostering a collaborative future for intelligent technologies.