Axioms for AI Alignment from Human Feedback
This work addresses AI alignment for researchers and practitioners by providing a new axiomatic framework to improve reward learning, though it is incremental as it builds on social choice theory.
The paper tackles the problem of learning reward functions in reinforcement learning from human feedback (RLHF) by framing it as a preference aggregation issue within social choice theory, and it demonstrates that existing models like Bradley-Terry-Luce fail basic axioms, leading to the development of novel rules with strong axiomatic guarantees.
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.