ASI-Bench removes human guidance mid-project and watches scores collapse from 50.91 to 26.62.

40+ experts and 31,000+ human hours built 60 project-level research tasks across 11 domains, designed so guidance is progressively withdrawn within the same project. Across 18 agent-model configurations: 50.91 with full methodological guidance, 29.10 when only the method is specified, 26.62 when the agent picks the method itself. (arXiv 2608.17271) Half the score was the human's methodology. Same conclusion as the Princeton study from a completely different direction.