Chen Wei

SoftmaxGRPO reallocates learning signals to improve performance on challenging prompts, achieving a 68.0% success rate on Poetry with minimal reward overhead.

Transforming diverse visual tasks into a unified RGB format enables a single model to excel across multiple domains without task-specific tuning.

Outlier tokens in Diffusion Transformers aren't just extreme values; they corrupt local patch semantics, and can be tamed with Dual-Stage Registers to boost image generation quality.

Time-frequency feature extraction via fractional Fourier transform unlocks surprisingly high-quality music generation from LSTMs.

LLMs can autonomously discover novel neural architectures that achieve state-of-the-art performance in specialized domains, suggesting a path towards automated scientific discovery.