ASI-Evolve: Towards AI-Driven Discovery Across the AI Research Stack

Can AI accelerate the development of AI itself? While recent agentic systems have shown strong performance on well-scoped tasks, it remains unclear whether they can tackle long-horizon research loops that drive real AI progress. In this paper, we present ASI-Evolve, an agentic framework for AI-for-AI research based on a learn–design–experiment–analyze cycle. ASI-Evolve delivers strong results across core areas of AI development: architectures, data, and algorithms. In neural architecture design, it discovers 105 SOTA linear attention architectures, with the best model surpassing DeltaNet by nearly 3× the gain of recent human-designed improvements. In pretraining data curation, its evolved data preparation pipeline yields training data that improves downstream model performance by +3.96 points on average, with gains exceeding 18 points on MMLU. In reinforcement learning algorithm design, it discovers algorithms that outperform GRPO by up to +12.5 points on AMC32, +11.67 points on AIME24, and +5.04 points on OlympiadBench. Further experiments in mathematics and biomedicine demonstrate the real-world effectiveness of this paradigm, showing that AI can iteratively improve both itself and its broader applications.