Axioms for AI Alignment from Human Feedback

TL;DR

This paper applies social choice theory to reinforcement learning from human feedback, proposing new reward aggregation rules with strong axiomatic guarantees and introducing the concept of linear social choice.

Contribution

It introduces novel reward aggregation rules with axiomatic guarantees and formulates a new paradigm called linear social choice for preference aggregation in AI alignment.

Findings

Existing models like Bradley-Terry-Luce fail basic axioms.

New rules meet key social choice axioms.

Linear social choice restricts feasible aggregation methods.

Abstract

In the context of reinforcement learning from human feedback (RLHF), the reward function is generally derived from maximum likelihood estimation of a random utility model based on pairwise comparisons made by humans. The problem of learning a reward function is one of preference aggregation that, we argue, largely falls within the scope of social choice theory. From this perspective, we can evaluate different aggregation methods via established axioms, examining whether these methods meet or fail well-known standards. We demonstrate that both the Bradley-Terry-Luce Model and its broad generalizations fail to meet basic axioms. In response, we develop novel rules for learning reward functions with strong axiomatic guarantees. A key innovation from the standpoint of social choice is that our problem has a linear structure, which greatly restricts the space of feasible rules and leads to a new paradigm that we call linear social choice.