Gen Li

Blurry structures and temporal flicker in driving scene rendering can be effectively mitigated using a unified framework that leverages spatial and temporal cues across datasets.

UCA-Flow achieves a remarkable 9.3% increase in action success rates while being up to 45.6 times faster than previous approaches.

Expert rerouting and attention-aware data packing can boost MoE model throughput by nearly 15%, reshaping efficiency in reinforcement learning.

$\omega$-0 enables humanoid robots to seamlessly integrate movement and manipulation, outperforming traditional models by predicting coordinated actions directly from sensory inputs.

ActiveVital achieves vital signs monitoring accuracy comparable to static methods by actively controlling radar alignment, reducing respiration interval error from 0.87 s to 0.14 s.

Gesture understanding can elevate robotic manipulation success rates by 80%, transforming how robots interpret human intentions.

RATrain achieves a remarkable 1.35x end-to-end speedup for LLM training on bandwidth-limited supercomputers, challenging the notion that high-bandwidth environments are necessary for efficiency.

PACT achieves state-of-the-art performance in medical dialogue systems by leveraging a unique combination of multi-paradigm synthesis and consensus training, all while safeguarding patient data.

Robots can now understand and act on your gaze while following language commands, enabling more intuitive and precise human-robot collaboration.

Robots can learn faster and perform better by strategically choosing *when* to be unpredictable, achieving both multimodal expressivity and deterministic efficiency.

Ditch brittle token rankings: OccamToken uses register-anchored relative evidence testing to prune visual tokens in VLMs, achieving extreme compression (down to 1.4%!) without retraining or significant accuracy loss.

Robots can now better understand implicit instructions in cluttered scenes, choosing the knife over the scissors when told to "cut the apple," thanks to a new benchmark and neural network architecture designed to handle functionally similar objects.