Toward Actionable Intelligence: Scaling Memory and Reasoning in AI Agents

Yangqin Jiang

The University of Hong Kong

2026年4月29日 星期三

下午 14:00-15:00

学院104会议室

Actionable intelligence for AI agents requires robust long-term memory, scalable reasoning, and reliable execution, yet it faces three core bottlenecks. In recommendation systems, structural memory—tasked with modeling user preferences—is often compromised by sparse interactions, irrelevant knowledge graph relations, and missing multimodal data, which can weaken preference representation. I address this by developing user preference memory frameworks that utilize adaptive graph learning, knowledge graph diffusion, and multimodal-aware modeling to refine and augment relational data. Regarding recommendation reasoning, current systems often rely heavily on ID-based embeddings, which can prioritize pattern memorization over genuine preference understanding and limit generalization. I aim to shift this reasoning paradigm toward semantic preference comprehension by leveraging LLM-powered foundation models to improve zero- and few-shot adaptability. Finally, agent actions in open-ended environments are often constrained by the "tool-use trilemma" of scale, quality, and granularity. I work to mitigate these hurdles by developing autonomous financial trading agents, lightweight on-device mobile foundation models, and a universal tool-use layer designed to unify tool scheduling. Collectively, these contributions strengthen the memory, reasoning, and execution capabilities of AI agents, supporting the development of more practical and effective actionable intelligence systems.