NeurIPS

“A multi-institution study used "shadow evaluation" — testing AI against unpublished NeurIPS 2026 submissions — and found that frontier AI agents could execute engineering tasks but failed to produce creative, original research contributions.”