Gemini 3.1 Powers AI Research Agents That Outperform Human Analysis

Gemini 3.1 cut hallucinations by 38 percentage points and hit 77.1% on ARC-AGI-2—but its research agents raise a uncomfortable question about human analysts.