A General Language Assistant as a Laboratory for Alignment

TL;DR

This paper explores methods to align large language models with human values, demonstrating that preference modeling outperforms imitation learning and that simple prompts can be effective without harming model performance.

Contribution

It introduces a comprehensive analysis of alignment techniques, showing the effectiveness of preference modeling and the benefits of simple prompting for large language models.

Findings

Preference modeling scales better than imitation learning.

Simple prompting benefits increase with model size.

Binary discrimination performs similarly to imitation learning.

Abstract

Given the broad capabilities of large language models, it should be possible to work towards a general-purpose, text-based assistant that is aligned with human values, meaning that it is helpful, honest, and harmless. As an initial foray in this direction we study simple baseline techniques and evaluations, such as prompting. We find that the benefits from modest interventions increase with model size, generalize to a variety of alignment evaluations, and do not compromise the performance of large models. Next we investigate scaling trends for several training objectives relevant to alignment, comparing imitation learning, binary discrimination, and ranked preference modeling. We find that ranked preference modeling performs much better than imitation learning, and often scales more favorably with model size. In contrast, binary discrimination typically performs and scales very similarly to imitation learning. Finally we study a `preference model pre-training' stage of training, with the goal of improving sample efficiency when finetuning on human preferences.