ARC-AGI-2

AI benchmark

A benchmark measuring fluid intelligence and novel abstract reasoning through visual grid puzzles. Models must identify patterns in input-output pairs and generate the correct output for unseen inputs. Considered the hardest public reasoning benchmark — average individual human performance is 66%.

Domains
Reasoning
Data source
benchlm-benchmarks
Catalogued
Website
https://benchlm.ai/benchmarks/arcAgi2

Catalogue entry from the RL Engineering daily scrape of public sources. Data reflects the 2026-08-30 snapshot.