Worth knowing
Updating the Epoch Capabilities Index methodology to use a Bayesian model
This methodological upgrade improves the statistical rigor of a prominent index used to aggregate benchmark scores and track the progress of frontier AI models.
What happened
Epoch AI has updated the methodology behind its Epoch Capabilities Index (ECI), transitioning from a frequentist approach to a Bayesian Item Response Theory model. This update replaces bootstrapping with Markov Chain Monte Carlo sampling and introduces a beta likelihood with benchmark-specific noise to provide more reliable uncertainty estimates and direct model comparisons.
Who it’s about
- Epoch AI
- Claude 3.5 Sonnet
- GPT-5
Source
More on benchmarks
Get a brief picked for you
Sign in to follow topics and companies, save stories and search the archive.
Sign in