Hanlin Gu

No existing LLM architecture can fully bridge the gap in graph inference tasks, with plain GNNs outperforming even the most advanced LLMs in critical areas like community detection.

Federated learning struggles when data quality varies across clients, but FedQual solves this with a novel approach that calibrates low-quality clients while preserving high-quality autonomy.

Byzantine-robust federated learning no longer needs to trade off convergence speed and model utility, even with a large number of malicious clients.

Dataset condensation, already vulnerable to backdoor attacks, now faces a far stealthier threat: InkDrop leverages decision boundary uncertainty to hide malicious triggers, making detection significantly harder.