Theoretical Limits of Language Model Alignment
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Researchers Lucas Monteiro Paes, Natalie Mackraz, Barry-John Theobald, and Federico Danieli derive a closed-form upper bound on reward gain for language model alignment under a fixed KL budget, governed by a Jeffreys divergence term.