Yongfeng Huang

Context construction can be dynamically tailored to specific queries, leading to substantial improvements in multi-hop question answering and summarization tasks.

You can now detect unauthorized training data in closed-source image generation models with higher accuracy, even without access to internal model features.

By generating synthetic gradients to bridge gaps between sparse benign updates, EnCAgg recovers more clean data in federated learning, even under dynamic model poisoning.

LLMs can now explore knowledge graphs on their own, discovering better reasoning paths and outperforming even closed-source models on question answering.