SuperARC benchmark says frontier models remain far from its AGI target

Researchers published SuperARC in Nature Communications as an open-ended benchmark grounded in algorithmic probability, compressed modeling, recursive prediction, and problem complexity. The authors designed it to resist saturation and contamination as conventional benchmarks become easier for frontier models to optimize against. Their evaluation finds leading systems relatively close to one another but far from the benchmark's proposed artificial general intelligence and artificial superintelligence targets. The paper explicitly cautions that a model could excel at SuperARC while lacking social intelligence, embodied reasoning, common sense, or goal-directed behavior, so the test should not be read as a complete measure of intelligence. Its value is methodological: it offers a new way to probe abstraction and prediction while being unusually clear about what benchmark performance cannot prove.

Key details: Published June 3, 2026, Nature Communications, Open-ended benchmark, Compressed modeling, Recursive prediction, Frontier models remain far from proposed AGI/ASI targets.

Continue swiping for more AI Brief stories.