ASI-Evolve: AI Accelerates AI

ASI-Evolve is an agentic AI-for-AI framework that closes the learn-design-experiment-analyze cycle and claims to be the first unified system demonstrating AI-driven discovery across data, architectures, and learning algorithms.

It augments evolutionary agents with a cognition base that injects accumulated human priors and includes a dedicated analyzer that distills complex experimental outcomes into reusable insights.

In neural architecture design it discovered 105 SOTA linear attention architectures, with the best model outperforming DeltaNet by +0.97 points.

In pretraining data curation the evolved pipeline improved average benchmark performance by +3.96 points and achieved MMLU gains exceeding 18 points.

In reinforcement learning algorithm design the discovered algorithms beat GRPO by up to +12.5 on AMC32, outperform AIME24 by +11.67, and OlympiadBench by +5.04; there is initial evidence the AI-for-AI paradigm can transfer to mathematics and biomedicine.