Jie Zhang

Fine-grained partitioning and targeted safeguards can provide robust NPU reliability with minimal performance overhead, challenging the assumption that redundancy is the only path forward.

BFP-based NPUs, despite their efficiency, exhibit surprising vulnerabilities to hardware faults that can be mitigated with a novel microarchitecture incurring minimal overhead.

Forget expensive labeled data: this method lets large reasoning models learn and outperform supervised RL by dynamically generating training data from unlabeled test queries.

A cleverly shaped and rendered adversarial patch can reliably spoof palmprint recognition, even across different models and datasets.