alignment.openai.comarticleBeyond g-mean²: complementary metrics for monitorability
Beyond g-mean² is an OpenAI alignment research post proposing complementary metrics for evaluating model monitorability. It says pair-matched comparisons and post-control risk should be used alongside g-mean². The fetched text is a single sentence, so the metric definitions, experiments and results are not available here.
TL;DR Proposes pair-matched comparisons and post-control risk as metrics that complement g-mean² for measuring model monitorability.
Organisations named here
- OpenAI
Techniques named here
- g-mean²
- Pair-matched comparisons
- Post-control risk
Related in evals
