Studies an analogy for supervising superhuman models: fine-tuning a strong model on labels produced by a much weaker one. Strong students often outperform their weak supervisors, but do not recover their full capability.
This is an external publication hosted at arxiv.org. AI Governance LK did not write it; the summary and notes on this page are our own.
Why it matters
The paper turns a hard-to-study future problem into an experiment that can be run today. It is a good model of how to design an analogy carefully and report its limitations honestly.
What you will learn
- Describe the weak-to-strong experimental setup
- Interpret “performance gap recovered” as a metric
Before you start
Suggested background. This is editorial guidance, not a requirement.
- Experience with fine-tuning
- Familiarity with scalable oversight
How to cite
Burns, C., Izmailov, P., Kirchner, J. H., et al. (2023). Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision. arXiv:2312.09390.