Mingyang Li

A systematic construction of compact, simply-connected, toric Einstein four-manifolds unlocks arbitrarily large Betti numbers and clarifies the topological landscape of these geometries.

A transferable residual adapter that injects additional ranking-specific features into ranking in a residual manner and an asymmetric multi-epoch training strategy that resets sparse parameters while continuously accumulating dense parameters across epochs, alleviating the overfitting of sparse parameters are proposed.

WLA$^3$ (World Latent Action Modeling for Semantics, Dynamics, and Kinematics), a unified generalist policy model framework built around representations learned by a World Latent Action Model (WLAM), is introduced.

SurfSVR transforms 2D surface information into robust 3D structures, eliminating common artifacts and achieving unprecedented reconstruction quality.

LOGOS shows that a single generative model can outperform specialized systems across a range of scientific tasks, challenging the need for separate AI frameworks in the natural sciences.

Real-robot trials are costly and slow, but DataLadder enables scalable evaluations and data generation through a seamless interplay between robots, simulations, and human feedback.

By explicitly aligning visuomotor representations with the action space's rigid-body geometry, OASIS achieves superior robotic manipulation performance compared to methods relying on implicit geometry recovery.

Bridging the gap between human manipulation and robotic control, JoyAI-RA unlocks enhanced cross-embodiment behavior learning through multi-source pretraining.