Tag
This paper analyzes test-time alignment methods that use a small aligned model as a proxy to guide a larger unaligned model's generation. The authors propose a novel rejection criterion based on conservative confidence betting and demonstrate improvements over existing approaches on multiple datasets.