A Fundamental Trade-off in Aligned Language Models and its Relation to Sampling Adaptors

Open in new window