Improving Weak-to-Strong Generalization with Reliability-Aware Alignment

TL;DR

This paper introduces a reliability-aware alignment method for large language models that improves their ability to generalize from imperfect supervision signals by estimating and utilizing the reliability of weak labels.

Contribution

The paper proposes a novel approach that incorporates answer reliability estimation into the alignment process to enhance weak-to-strong generalization in LLMs.

Findings

Effective identification of weak label quality

Significant improvement in generalization performance

Enhanced robustness to noisy supervision

Abstract

Large language models (LLMs) are now rapidly advancing and surpassing human abilities on many natural language tasks. However, aligning these super-human LLMs with human knowledge remains challenging because the supervision signals from human annotators may be wrong. This issue, known as the "super-alignment" problem, requires enhancing weak-to-strong generalization, where a strong LLM must generalize from imperfect supervision provided by a weaker source. To address this issue, we propose an approach to improve weak-to-strong generalization by involving the reliability of weak supervision signals in the alignment process. In our method, we query the weak supervisor for multiple answers, estimate the answer reliability, and enhance the alignment process by filtering out uncertain data or re-weighting reliable data. Experiments on four datasets demonstrate that our methods effectively identify the quality of weak labels and significantly enhance weak-to-strong generalization. Our work presents effective techniques for error-robust model alignment, reducing error propagation from noisy supervision and enhancing the accuracy and reliability of LLMs. Codes are publicly available at http://github.com/Irenehere/ReliableAlignment.