About Me
I am Weihan Li (黎维瀚), a PhD student at Georgia Tech CSE, advised by Prof. Anqi Wu. Previously, I was an undergraduate at Zhejiang University, working with Prof. Gang Pan. My CV is available here.
Research
I work on robot learning for dexterous manipulation: teaching multi-fingered robot hands to perform bimanual manipulation skills learned from human demonstrations, in simulation and on real hardware.
Before moving to robotics, I worked on modeling communication across multiple brain regions and on analyzing animal behavior from pose and video.
Current Dexterous Manipulation Bimanual Robot Skills Learning from Human Demonstrations Sim-to-Real Transfer
Previously Multi-Region Brain Communication Animal Behavior Analysis
Publications
See also my Google Scholar profile.
Learning When Visual Context Matters for Mouse Behavior Analysis
NeurIPS 2026
Frames with large quantization residuals in a pose-based motion encoder are where visual context helps most; ResiFuse adds video features only on those frames and outperforms pose- and video-based baselines on most tasks across three mouse behavior datasets.
MICE: Multi-animal Interaction Context Encoder — A Hierarchical Foundation Model for Mouse Behavior
NeurIPS 2026Spotlight
A hierarchical self-supervised foundation model for mouse behavior, pretrained jointly on six mouse-pose datasets with one to four animals; a Perceiver-style social encoder lets a single checkpoint handle any group size, matching or exceeding prior keypoint-based baselines.
QDMouse4M: A Multi-View 3D Mouse Spontaneous Behavior Dataset with Quantum-Dot Markers
NeurIPS 2026 · Evaluations and Datasets Track
A six-view dataset of freely moving mice with over four million frames, whose 2D and 3D keypoint labels come from subdermal quantum-dot fluorescence markers; used to benchmark markerless 2D/3D pose estimation and downstream spontaneous-behavior analysis.
BehaviorVLM: Unified Finetuning-Free Behavioral Understanding with Vision-Language Reasoning
Preprint
A unified framework for animal pose estimation and behavioral understanding that guides pretrained vision-language models through explicit, verifiable reasoning steps, with no task-specific finetuning and minimal human labeling.
A Factorized Low-Rank RNN Framework for Uncovering Independent Neural Latent Dynamics and Connectivity
ICML 2026Spotlight
FacRNN, a generative low-rank RNN framework that encourages group-wise independence among latent dynamics, yielding more disentangled and interpretable latent trajectories and low-rank connectivity.
Uncovering Semantic Selectivity of Latent Groups in Higher Visual Cortex with Mutual Information-Guided Diffusion
ICLR 2026
MIG-Vis infers group-wise disentangled neural latent subspaces from macaque IT cortex and uses mutual-information-guided diffusion to visualize the visual-semantic features each latent group encodes.
Learning Time-Varying Multi-Region Brain Communications via Scalable Markovian Gaussian Processes
ICML 2025Oral
The Adaptive Delay Model combines Gaussian processes with state-space models to learn multi-region brain communication with time-varying delays, using parallel-scan inference to scale to long recordings.
Multi-Region Markovian Gaussian Process: An Efficient Method to Discover Directional Communications Across Multiple Brain Regions
ICML 2024
Builds a linear dynamical system that mirrors a multi-output Gaussian process with explicit frequencies and phase delays, revealing communication directions across brain regions and separating oscillatory communication into frequency bands at linear cost.
A Differentiable Partially Observable Generalized Linear Model with Forward-Backward Message Passing
ICML 2024
A differentiable partially observable GLM that enables pathwise gradients, paired with a forward–backward message-passing variational model, for learning neural connectivity in the presence of hidden neurons.
Efficient Point-Process Modeling of Spiking Neurons for Neuroprosthesis
EMBC 2021
A spiking neuron point-process model that takes spike times as input to model nonlinear spike transformations between cortical areas (PMd to M1), and can run on low-energy neuromorphic hardware for neuroprostheses.
Service
Reviewer: NeurIPS, ICML, ICLR