Xin Liu

Noisy multimodal preference datasets are holding back reward model performance, but DT2IT-MRM offers a scalable curation strategy that achieves state-of-the-art results.

FLASH enables robots to master complex deformable manipulation tasks in minutes using only synthetic data, eliminating the need for labor-intensive real-world training.

Reasoning with LLMs just got a whole lot faster: MemoSight cuts KV cache footprint by 66% and speeds up inference by 1.56x without sacrificing CoT performance.

Forget painstakingly teleoperating robots or waiting for CPU planners: AutoMoMa unlocks 80x faster generation of kinematically valid mobile manipulation trajectories, finally making large-scale data practical.

Mimicking the human visual system's "glance and gaze" strategy lets video transformers capture long-range dependencies and achieve state-of-the-art action recognition.