FRAMES: leaderboard

824 multi-hop questions each needing 2 to 15 Wikipedia articles, tagged by reasoning type (numerical, tabular, temporal, constraints), from Google DeepMind and Harvard (2024); LLM-graded accuracy.

Source: arxiv.org.

Interactive version: theaggregate.ai/benchmark?slug=frames · How It Works · Data refreshed daily, snapshot 2026-09-05.