Uses language models to generate large evaluation datasets cheaply, then uses them to surface behaviours such as sycophancy that become more pronounced with scale or with RLHF.
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
Generating evaluation data with models makes it feasible to cover many behaviours and languages, which is directly relevant to work in under-resourced languages. The paper’s quality-checking steps are as important as its findings.
What you will learn
- Describe a pipeline for generating and filtering evaluation items
- Explain the risks of models evaluating models
Before you start
Suggested background. This is editorial guidance, not a requirement.
- Basic understanding of benchmarks
How to cite
Perez, E., Ringer, S., Lukošiūtė, K., et al. (2022). Discovering Language Model Behaviors with Model-Written Evaluations. arXiv:2212.09251.