Maximizing the efficiency of human feedback in AI alignment: a comparative analysis

TL;DR

This paper investigates alternative sampling strategies for preference inference in RLHF, introducing Swiss InfoGain, which improves efficiency and robustness in human feedback utilization for AI alignment.

Contribution

It proposes Swiss InfoGain, a novel adaptive sampling method inspired by game theory, that outperforms traditional Bradley-Terry sampling in constrained annotation settings.

Findings

Swiss InfoGain significantly outperforms Bradley-Terry in sample efficiency.

Adaptive strategies reduce redundancy and improve robustness.

Preference inference quality improves with resource-aware sampling.

Abstract

Reinforcement Learning from Human Feedback (RLHF) relies on preference modeling to align machine learning systems with human values, yet the popular approach of random pair sampling with Bradley-Terry modeling is statistically limited and inefficient under constrained annotation budgets. In this work, we explore alternative sampling and evaluation strategies for preference inference in RLHF, drawing inspiration from areas such as game theory, statistics, and social choice theory. Our best-performing method, Swiss InfoGain, employs a Swiss tournament system with a proxy mutual-information-gain pairing rule, which significantly outperforms all other methods in constrained annotation budgets while also being more sample-efficient. Even in high-resource settings, we can identify superior alternatives to the Bradley-Terry baseline. Our experiments demonstrate that adaptive, resource-aware strategies reduce redundancy, enhance robustness, and yield statistically significant improvements in preference learning, highlighting the importance of balancing alignment quality with human workload in RLHF pipelines.