Concept计算最优缩放 / 推理时计算缩放 / Test-Time Compute Scaling
Compute-Optimal Scaling
A research area in machine learning concerned with optimal allocation of training and inference compute resources. It encompasses training-time scaling laws identified in works such as the Chinchilla paper, as well as methods for dynamically increasing inference-time computation to improve model output quality.
Timeline (last 90 days)
Oct 6
DeepMind's Chinchilla paper and subsequent research on inference-time compute scaling show that dynamically increasing computation during the inference phase can significantly improve the output quality of smaller models.
Unverified50%