Grounding
Anchor concepts in the current world: its state, affordances, physical constraints, and causes.
Exploring the world
Zhiwei Yu.
I design smart systems—digital or embodied—to turn real-world chaos into innovation—where AI, product, and business collide. My research interests span World Modeling, Embodied AI, and Knowledge Computing.
Previously, I graduated from AAIS and WICT at Peking University, advised by Xiaojun Wan. Then I joined MSRA as a researcher working in Chin-Yew Lin's team. During my stay in MS, I drove research on Knowledge Base and steered AI integration across flagship products such as Office, Bing and Windows.
Senior Researcher
Multimodal Interaction Research Center
Beijing Academy of Artificial Intelligence
A research practice for dynamic worlds
01 / Research interests
Intelligent Interaction in Dynamics
Evolving Intelligence for an Ever-Changing World.
Real environments do not hold still. People differ, conditions shift, feedback arrives late, and a once-valid plan can become unsafe or infeasible.
My research focuses on systems that take verifiable actions under uncertainty and continually revise themselves through interaction.
Experience
Turn feedback into updated beliefs, useful memory, and skills that can be reused beyond one task.
Adaptation
Make experience persist across tasks, time, and individuals without resetting the intelligent system.
01 · Lifelong Agents
02 · World Models & Memory
03 · Human–AI Symbiosis
Lifelong Agents
Benchmark · Closed-loop systems · Recovery
Long-horizon interaction requires more than completing a prescribed sequence. We study how agents perceive changing state, verify consequences, recover from failure, and turn execution experience into supervision—from diagnostic benchmarks to systems that can improve over time.
SWITCH
v0 · ICLR 2026 Lifelong Agents v1 · EMNLP 2026Benchmarking modeling and handling of tangible interfaces in long-horizon embodied scenarios.
From Failures to Supervision
Under reviewA closed-loop planning system that learns robust recovery from dynamic execution failures.
World Models & Memory
World state · Structured knowledge · Long-term memory
Intelligent systems need representations that preserve what has changed, explain why it changed, and retrieve the knowledge that matters now. This theme connects semantically rich world models, memory roles, and retrieval-augmented reasoning across physical and digital environments.
Semantically Rich World Models
Under review · ICLR 2027World representations that expose meaningful state and interaction structure for physical planning.
Memory Makes the Difference
EMNLP 2026 · GLM WorkshopEvaluating how distinct memory roles shape behavior, accuracy, and personalization in agents.
TIARA
EMNLP 2022Multi-grained retrieval for robust reasoning and generalization over large knowledge bases.
ReTraCk
ACL-IJCNLP 2021 · DemoA modular retriever–transducer–checker framework for efficient knowledge-base reasoning.
Human–AI Symbiosis
Human signals · Machine humor · Human-centered interaction
Human–AI symbiosis connects how machines act with how they communicate. We explore human signals that guide embodied behavior, alongside humor and creative language that make digital interaction more expressive and engaging. Together, these directions bring human intent, experience, and interpretation into the design of intelligent systems.
NeuroEmbody
OngoingBrain, muscle, vision, and behavior signals that connect human feedback to perception, decisions, and autonomous action in the physical world.
Machine Humor & Creative Language
Generating puns through ambiguity and semantic incongruity, and exploring metaphor as a related form of creative expression—toward richer communication between people and machines.