Zhuosheng Zhang

MLLMs excel at single-hop tasks but falter dramatically in open-world scenarios, revealing critical gaps in their reasoning capabilities.

GUI agents can learn world knowledge more efficiently by internalizing causal relationships during mid-training, rather than relying on implicit learning through action annotations or reward signals in post-training.

Mobile-using agents can now achieve 17% higher task success rates offline and 26% higher success rates in real-world dynamic experiments, all while significantly reducing unnecessary human intervention.

Poisoned training data leaves a unique fingerprint in the spectral entropy of LLM gradients, enabling backdoor detection even at extreme poison ratios where clustering-based defenses fail.

Pinpoint exactly which client leaked your federated model with a black-box watermark that's robust to fine-tuning, pruning, and quantization.