AI Evaluation and Benchmarking

Overview

This area tracks how model capabilities are measured, how benchmarks saturate or fail, and which metrics are needed to say something useful about model performance rather than just leaderboard position.

Active Questions

  • How do we design benchmarks that stay informative as frontier systems rapidly improve?
  • Which probabilistic metrics should accompany accuracy so that confidence quality and monitoring reliability are visible?
  • How do public benchmarks resist contamination, overfitting, and metric gaming?
  • Can reasoning traces be monitored without optimizing away their informativeness?
  • When do benchmark or capability curves saturate, and when can model/data scaling produce nonmonotone generalization effects?
  • When do implementation choices make an algorithm look weak or strong on a benchmark independent of the core idea being tested?
  • Which cooperative-MARL baselines are necessary before claiming progress in safety, coordination, or sample efficiency?
  • Which MARL benchmark matrices include enough task diversity, observation modalities, variance reporting, and training-time cost to support general claims?
  • Where do empirical evaluations stop being enough, requiring model-based quantitative safety guarantees?

Key Concepts

Key Sources

Adjacent Foundations