Traditional approaches to improving AI model performance—scaling model size or training data—are increasingly constrained by cost, latency, and diminishing returns.

Inference-Time Scaling (ITS) offers an orthogonal solution by optimizing how computational resources are allocated during inference.By restructuring search and evaluation strategies at test-time, ITS significantly enhances model output quality without retraining or expanding model parameters.

In this talk, we will introduce the top methods of ITS, and how you can try it on your existing models using off-the-shelf tool-kits such as reward_hub and inference_time_scaling library.