A General Language Assistant as a Laboratory for Alignment

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.

26+19 pages; v2 typos fixed, refs added, figure scale / colors fixed; v3 correct very non-standard TruthfulQA formatting and metric, alignment implications slightly improved

References in corpus (9)

• Language Models are Few-Shot Learners
• Scaling Laws for Neural Language Models
• The Pile: An 800GB Dataset of Diverse Text for Language Modeling
• Generating Wikipedia by Summarizing Long Sequences
• Imitating Interactive Intelligence
• HellaSwag: Can a Machine Really Finish Your Sentence?
• Program Synthesis with Large Language Models
• Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics
• Scaling Scaling Laws with Board Games