Inference-time compute has emerged as a new axis for scaling large language models, leading to breakthroughs in AI reasoning. Broadly speaking, inference-time compute methods involve allowing the language model to interact with a verifier to search for desirable, high-quality, or correct responses. While recent breakthroughs involve using a ground-truth verifier of correctness, it is also possible to invoke the language model itself or an otherwise learned model as verifiers. These latter protocols raise the possibility of self-improvement, whereby the AI system evaluates and refines its own generations to achieve higher performance.

This talk presents new understanding of and new algorithms for language model self-improvement. The first part of the talk focuses on a new perspective on self-improvement that we refer to as sharpening, whereby we "sharpen" the model toward one placing large probability mass on high-quality sequence, as measured by the language model itself. We show how the sharpening process can be done purely at inference time or amortized into the model via post-training, thereby avoiding expensive inference-time computation. In the second part of the talk, we consider the more general setting of a learned reward model, show that the performance of naive-but-widely-used inference-time compute strategies does not improve monotonically with compute, and develop a new compute-monotone algorithm with optimal statistical performance.

Based on joint works with Audrey Huang, Dhruv Rohatgi, Adam Block, Qinghua Liu, Jordan T. Ash, Cyril Zhang, Max Simchowitz, Dylan J. Foster and Nan Jiang.