Measuring Progress on Scalable Oversight for Large Language Models
This addresses the challenge of supervising AI systems that may outperform humans, with incremental progress in empirical study methods.
The paper tackles the problem of scalable oversight for AI systems by proposing an experimental design using tasks where human specialists succeed but unaided humans and current AI fail, and demonstrates that humans interacting with an unreliable large language model assistant substantially outperform both the model alone and their unaided performance on MMLU and time-limited QuALITY tasks.
Developing safe and useful general-purpose AI systems will require us to make progress on scalable oversight: the problem of supervising systems that potentially outperform us on most skills relevant to the task at hand. Empirical work on this problem is not straightforward, since we do not yet have systems that broadly exceed our abilities. This paper discusses one of the major ways we think about this problem, with a focus on ways it can be studied empirically. We first present an experimental design centered on tasks for which human specialists succeed but unaided humans and current general AI systems fail. We then present a proof-of-concept experiment meant to demonstrate a key feature of this experimental design and show its viability with two question-answering tasks: MMLU and time-limited QuALITY. On these tasks, we find that human participants who interact with an unreliable large-language-model dialog assistant through chat -- a trivial baseline strategy for scalable oversight -- substantially outperform both the model alone and their own unaided performance. These results are an encouraging sign that scalable oversight will be tractable to study with present models and bolster recent findings that large language models can productively assist humans with difficult tasks.