Pengfei Zhang

DreamX-World 1.0 not only generates high-quality videos but also allows for precise camera navigation and event control, setting a new benchmark in interactive world modeling.

AP-GRPO reveals that leveraging reliable speech anchors can significantly enhance the reconstruction of distorted speech, adapting to the severity of neurodegenerative conditions.

A small fraction of sharp dimensions governs the generalization of PEFT methods, and optimizing these can lead to substantial performance gains.

Reward-guided optimization can be selectively applied to enhance generative recommendation performance, avoiding the pitfalls of uniform reinforcement learning.

Skewed item distributions in recommendation systems can be tamed with a learnable non-uniform quantization, leading to better codebook utilization and more accurate generative recommendations.

Forget blindly aligning layers in audio Flow Matching: AG-REPA reveals that targeting the *causally dominant* layers driving the velocity field, not just representationally rich ones, unlocks superior generation.

Generative recommendation can beat DLRM in large-scale advertising, driving a 4.2% revenue lift in Kuaishou's production system via innovations in tokenization, decoding, optimization, and serving.