Yuan Zhang

TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference

Tinghao Wang, Yichen Guo, Rui Huang, et al.

DiT-Reward: Generative Representations for Text-to-Image Reward Modeling

Yuanming Yang, Guoqing Ma, Bo Wang, et al.

Generalized Priority-Aware Shapley Value

Kiljae Lee, Ziqi Liu, Weijing Tang, et al.

First-Order Efficiency for Probabilistic Value Estimation via A Statistical Viewpoint

Ziqi Liu, Kiljae Lee, Yuan Zhang, et al.

Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising

Jun Guo, Qiwei Li, Peiyan Li, et al.

SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits

Edward Lin, Sahil Modi, Siva Kumar Sastry Hari, et al.

SOL-ExecBench is a new benchmark for evaluating AI systems that optimize GPU kernels by measuring how close they get to the theoretical maximum performance (Speed-of-Light) that the hardware can achieve, rather than just comparing against other software implementations. It includes 235 real optimization problems from actual AI models and uses a special pipeline called SOLAR to calculate what perfect performance would look like for each kernel. The benchmark includes safeguards to prevent cheating and focuses on modern NVIDIA Blackwell GPUs with various precision formats.