Hao Jiang

AVOC achieves a remarkable 4.9-point accuracy boost over the next best model in long-form audio-video comprehension, redefining efficiency in multimodal understanding.

Surprisingly, the "think before answer" paradigm fails to enhance generative recommendation models, prompting a novel approach that redefines how reasoning is integrated into these systems.

LoomVideo achieves state-of-the-art video generation and editing efficiency with a compact architecture that accelerates inference speed by over 5 times compared to larger models.

Online RL for LLMs suffers from an imbalanced exploration-exploitation trade-off, but IB-TPO solves this with a novel Information Bottleneck-driven tree sampling strategy that boosts performance by up to 3.6%.

Stop wasting compute on redundant retrieval calls: DynFrame learns the optimal frame sampling density *within* each temporal window, slashing context length and boosting performance on complex video understanding tasks.

Even slight tracking errors can drastically degrade personal sound zone performance, but a simple neighbor-consistency regularization can dramatically improve robustness.

Prioritizing resource-aware grasps, where finger usage is explicitly modeled, dramatically improves the success rate of sequential dexterous manipulation tasks.

Skip the costly human annotations: PromptEcho distills reward signals directly from frozen VLMs to boost text-to-image RL, achieving state-of-the-art results without any reward model training.

Forget end-to-end training: DexMulti's "retrieve-align-execute" approach lets robots master complex, multi-stage dexterous tasks from just a handful of demonstrations.