Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner.

Bolian Li, Yanran Wu, Xinyu Luo, Ruqi Zhang

Browse the full EMNLP paper archive.