Zhen Li

Z-Reward achieves nearly 90% human preference accuracy by transforming subjective visual preferences into nuanced score distributions, outperforming traditional reward models.

Resource-constrained edge devices can achieve Pareto-optimal trade-offs between inference accuracy, latency, and energy consumption in federated learning by using a constrained multi-objective reinforcement learning approach.

Fine-tuning efficient few-step diffusion models no longer requires sacrificing their speed, thanks to a self-distillation approach that preserves inference capabilities.

Forget reward function dependencies – this new approach to contextual bandits with latent state dynamics achieves stronger regret bounds by directly modeling hidden state dependencies and adaptively estimating HMM parameters.

Pre-training with Dual Latent World Models unlocks significant performance gains in autonomous driving tasks by learning holistic Gaussian-centric representations.