Hong Xu

AIR achieves unprecedented attack success rates against subject-agnostic face swapping models while maintaining high visual fidelity, without needing a surrogate model.

Achieving up to 3.0x higher throughput in agentic RL, Libra redefines resource management for large language models under complex workloads.

Unconstrained reflection loops in multi-agent systems actively induce logical drift, challenging the assumption that more reflection always leads to better reasoning.