gapnium

← Timeline · Benchmark

Humanity's Last Exam: models below 10%

· significance ●●○○○

A 3,000-question expert-written benchmark finds frontier models scoring under 10%, with poor calibration.

What happened and why it matters

Created by the Center for AI Safety and Scale AI with nearly 1,000 subject experts, HLE was built to be the hardest closed-ended academic benchmark. Low scores and high overconfidence quantified the remaining knowledge-and-reasoning gap.

Sources

Hover the chart to read values · click a marker to open its articles