A density estimation perspective on learning from pairwise human preferences
This work addresses a crucial challenge in aligning AI systems with human values, though it is incremental by offering an alternative theoretical perspective rather than a new method.
The paper tackles the problem of learning from pairwise human preferences in large language models by reframing it as a density estimation problem, showing that training a reward function effectively models annotators' implicit preference distributions and identifying failure cases due to annotator misspecification.
Learning from human feedback (LHF) -- and in particular learning from pairwise preferences -- has recently become a crucial ingredient in training large language models (LLMs), and has been the subject of much research. Most recent works frame it as a reinforcement learning problem, where a reward function is learned from pairwise preference data and the LLM is treated as a policy which is adapted to maximize the rewards, often under additional regularization constraints. We propose an alternative interpretation which centers on the generative process for pairwise preferences and treats LHF as a density estimation problem. We provide theoretical and empirical results showing that for a family of generative processes defined via preference behavior distribution equations, training a reward function on pairwise preferences effectively models an annotator's implicit preference distribution. Finally, we discuss and present findings on "annotator misspecification" -- failure cases where wrong modeling assumptions are made about annotator behavior, resulting in poorly-adapted models -- suggesting that approaches that learn from pairwise human preferences could have trouble learning from a population of annotators with diverse viewpoints.