<ul><li>Cascade Reward Sampling (CARDS) is introduced to address efficiency bottlenecks in decoding-time alignment of large language models (LLMs).</li><li>CARDS utilizes a segment-level rejection sampling algorithm to minimize redundant computations of LLMs and reward models (RMs).</li><li>An uncertainty-based segmentation mechanism ensures accurate evaluation of RMs on incomplete segments.</li><li>Experimental results demonstrate that CARDS significantly improves decoding efficiency, alignment quality, and general utility.</li></ul>

Cascade Reward Sampling for Efficient Decoding-Time Alignment

Discover more