Jianbiao Mei

I received my Ph.D. in June 2026 from the Department of Control Science and Engineering at Zhejiang University, advised by Prof. Yong Liu at the APRIL Lab. Before that, I received my B.Eng. in June 2021 from the same department with an honors degree at Chu Kochen Honors College.

I am currently a researcher at Shanghai AI Laboratory. My research broadly focuses on embodied and interactive AI, including autonomous agents, closed-loop environments, multimodal reasoning, and interactive systems.

Email  /  Scholar  /  GitHub

News

Research

* indicates equal contributions, † indicates equal advising.

Context-conditioned reconstruction turns retrieved agent memories into state-aligned guidance instead of static replay.

Proxy-guided update signal transfer decouples low-cost exploration from reusable policy alignment for LLM post-training.

Anchoring user intent and verifying evidence locally enables robust multi-turn search with LLMs.

Training search agents with reinforcement learning enables open-domain open-ended question answering.

Sequential planning and iterative reflection enable self-evolving action-conditioned video generation for long-horizon tasks.

Adaptive switching between strategic planning and reactive execution improves long-horizon game agents under token and latency budgets.

Distilling and reusing interaction experience enables self-evolving LLM agents.

Simulating first-day workplace scenarios enables TraineeBench to evaluate agent learning, exploration, and scheduling.

Accumulating hierarchical experience enables MUSE to self-evolve on long-horizon productivity tasks.

Modeling residual scene changes enables vision-centric 4D occupancy forecasting and planning.

Evaluating multi-hop reasoning over ultra-high-resolution imagery reveals the limits of current multimodal models.

Constrained trajectory reasoning over knowledge graphs improves attribution-aware large language model reasoning.

Generating 3D scenes from semantic and textual controls enables high-fidelity controllable driving scene synthesis.

Organizing video, occupancy, and LiDAR generation methods clarifies the landscape of 3D and 4D world modeling.

Selecting goal-relevant objects with dual-process reasoning enables demand-driven robot navigation.

Coupling generative scene simulation with closed-loop traffic interaction enables realistic evaluation of driving agents.

Autoregressively generating motion-aware multi-view videos enables long-horizon controllable driving scene synthesis.

Decoupling thing and stuff queries enables unified LiDAR panoptic segmentation.

Forecasting controllable 4D occupancy with world models enables vision-centric planning for autonomous driving.

Combining cognitive perception with dual-process thinking improves autonomous driving decision making.

Combining dual-process reasoning with experience memory enables continuously improving autonomous driving agents.

Propagating sparse semantic and geometric guidance enables camera-based 3D semantic scene completion.

Distilling LiDAR-camera occupancy knowledge into radar inputs improves robust semantic occupancy prediction.

Combining attention-guided descriptors with overlap estimation enables coarse-to-fine place recognition.

Storing short-term and long-term point-cloud memory improves 3D single object tracking under sparse observations.

Adapting multimodal CLIP features with multi-task supervision improves video action recognition and transfer.

Matching satellite imagery with LiDAR point clouds enables accurate geo-localization for urban navigation.

Representing LiDAR instances by sparse centers enables detection-free and clustering-free panoptic segmentation.

Separating semantic and geometric representations with BEV fusion improves LiDAR semantic scene completion.

Generating sparse instance proposals and aggregating fragments enables efficient LiDAR panoptic segmentation.

Organizing compact mask-aware memory enables fast real-time video object segmentation.

Adapting language-image pretrained models enables transferable video action recognition across supervised and zero-shot settings.

Disentangling spatial and temporal context accelerates neural video representation with fewer redundant parameters.

Exploiting reference masks more effectively improves matching and boundary reasoning for video object segmentation.

Modeling spatiotemporal relationships with transformers enables a unified pipeline for video object segmentation.

Service

Reviewer of TPAMI, TIP, TMM, TCSVT, RAL, CVPR, ECCV, NeurIPS, ICLR, AAAI, ACM MM, BMVC, ICRA, IROS, et al.

This website is adapted from this amazing template.