Hao Tang

ReMoMask-2, a structure-aware RAG framework coupling Hierarchical Bidirectional Momentum contrastive learning to align global and part-level features with text; Semantic Spatial-Temporal Attention (SSTA) for topology-aware fusion; and Topology Structured Masking (TSM) to force robust part-level grounding via adaptive masking are introduced.

This work proposes MobileVLA-R1 2.0, an RL-enhanced VLA framework that explicitly couples structured embodied reasoning with executable mobile robot control, and introduces a reasoning-conditioned action decoder that maps multimodal reasoning representations to task-level action targets, which are subsequently translated into embodiment-specific commands by robot controllers.

Achieving coupled-cluster accuracy for molecular properties with the computational cost of a single DFT calculation could revolutionize molecular modeling in chemistry.

Geometry-consistency awareness in video spatial reasoning can lead to a 12.6-point performance boost over current state-of-the-art models.

Geometry-aware reconstruction reduces hallucinations in object poses, leading to more reliable robot planning.

Factorized Neural Operators achieve superior physical modeling by separating transient and persistent responses, leading to enhanced accuracy and interpretability.

A dual-stream tokenizer reveals that traditional methods misrepresent high-frequency motion dynamics, leading to over 50% improvement in diversity alignment with real data.

DragMesh-2 achieves unprecedented robustness in dexterous manipulation of articulated objects, outperforming traditional methods even under varying contact loads.

Encoding scene graphs in hyperbolic space unlocks significantly better hierarchical structure quality compared to Euclidean embeddings, leading to state-of-the-art graph-level scene understanding.

Forget fragmented architectures: UniMesh lets you edit 3D meshes iteratively with semantic prompts, bridging the gap between generation and understanding.

MLLMs struggle with spatial reasoning because they flatten geometric information too early; GUIDE fixes this by progressively injecting multi-granularity geometric priors into early layers, leading to substantial performance gains.

By explicitly enforcing action-conditioned consistency during training and distillation, MWM enables more reliable planning in imagined future spaces for embodied navigation.

By swapping out ConvGRUs for a novel ConvSS2D operator, StereoAdapter-2 achieves state-of-the-art zero-shot underwater stereo depth estimation with a single update step.

Achieve zero-shot robotic manipulation by guiding a hierarchical vision-language-action model with knowledge-guided trajectory planning, outperforming existing methods and eliminating the need for real-world data collection.