Can AI research itself without losing human oversight?

Recent agentic systems perform well on well-scoped tasks with rapid feedback, but the loops that actually drive AI progress are costly, long-horizon, and weakly supervised. ASI-Evolve targets that regime with a learn–design–experiment–analyze cycle, and its contribution is what it adds to a standard evolutionary agent: a cognition base that injects accumulated human priors into each round of exploration, and a dedicated analyzer that distills complex experimental outcomes into reusable insights for future iterations.

The framing worth keeping is diagnostic. The paper locates the bottleneck of AI research not in raw search but in three human constraints: the hypothesis space humans can explore in parallel is small, experimental workflows demand constant manual intervention, and insights accumulated across iterations depend on individual intuition and so transfer poorly. The two added components map directly onto the last two — the analyzer makes insight accumulation systematic rather than intuitive, and the cognition base substitutes for the priors a human would carry between experiments.

The empirical reach is the claim's evidence: AI-driven discovery across all three central components of AI development — data curation (+3.96 avg, >18 on MMLU), neural architecture (105 SOTA linear-attention designs), and learning algorithms. This complicates the autonomy debate. Where Can human-AI research teams improve faster than autonomous AI systems? argues for keeping humans in the loop, ASI-Evolve shows the priors and insight-distillation a human normally supplies can themselves be made into loop components — which is exactly what makes the autonomy question urgent.