Zilong Huang

Emotion preference models can be dramatically improved by addressing both data sparsity and model bias, leading to more accurate emotional assessments in multimodal contexts.

Replacing standard residual connections with manifold-constrained hyper-connections boosts speaker recognition performance across multiple architectures.

Speaker recognition accuracy improves dramatically when leveraging a U-Net-based fusion of noisy and enhanced speech, coupled with a novel training strategy.

LLMs can now scale depth more effectively: a new attention mechanism recovers diluted features in deeper layers, boosting performance with negligible overhead.

Achieve state-of-the-art video generation with 24% fewer tokens by adaptively allocating tokens based on video segment complexity.