A General Language Assistant as a Laboratory for Alignment
This work addresses the critical challenge of ensuring AI safety and alignment for general-purpose language models, though it is incremental as it builds on existing techniques like prompting and preference modeling.
The paper tackles the problem of aligning large language models with human values to create helpful, honest, and harmless assistants, finding that ranked preference modeling outperforms imitation learning and scales better with model size, with modest interventions improving alignment without compromising performance.
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.