Tag
The paper presents a theoretical framework for steering and scaling large language models via sampling algorithms, such as Sequential Monte Carlo and Replica Exchange, to improve generation quality without external supervision.