Yining Li

OPDVR transforms the landscape of model distillation by ensuring that only correct trajectories enhance learning, leading to significant performance gains on reasoning tasks.

Reciprocal cross-stream addressing in Dual Attention Residuals leads to significant improvements in Transformer performance without sacrificing depth-wise diversity.

Achieving comparable text-to-image quality with a linearized model that accelerates inference by up to 1.47 times, all while leveraging pretrained weights.

A stark capability cliff reveals that even leading AI models falter on complex workflows, achieving less than 15% success despite advancements in tool-use benchmarks.

Automating LLM fine-tuning is now possible: a multi-agent system, TREX, matches or exceeds human performance on a diverse set of real-world tasks.

LLMs can now automatically evolve and optimize GPU kernels to beat hand-tuned and proprietary models like Gemini and Claude.

Optimism is the key to stable and convergent safe RLHF, according to a new primal-dual framework that unifies existing alignment algorithms.