ASI-Evolve: When AI Designs Its Own Architecture, Humans Can't Keep Up
A team from Shanghai Jiao Tong University's GAIR group just released ASI-Evolve, an agentic framework where AI autonomously researches, designs, experiments, and analyzes to improve itself. Not as a concept paper — as working code with results that make human researchers look slow.
The framework runs a tight loop: a Researcher Agent proposes candidates using domain knowledge, an Engineer Agent builds and benchmarks them, an Analyzer Agent extracts reusable insights, and a Cognition Store accumulates all learned priors for the next iteration. Think of it as autoresearch with memory — every experiment makes the next one smarter.
The numbers are hard to argue with. In neural architecture search, ASI-Evolve discovered 105 SOTA linear attention architectures. The best one beat DeltaNet by 0.97 points — nearly 3x the gain of recent human-designed improvements. In data curation, the evolved pipeline improved benchmark performance by 3.96 points average, with MMLU gains exceeding 18 points. In RL algorithm design, discovered algorithms outperformed GRPO by up to 12.5 points on AMC32.
What makes this different from prior autoresearch is the Cognition Store. Instead of starting fresh each run, accumulated human priors and experimental insights carry forward. The system doesn't just search — it learns what kinds of architectures work and why, then uses that understanding to guide future exploration.
The code is fully open-source under Apache 2.0 at github.com/GAIR-NLP/ASI-Evolve. If the trend of AI designing AI holds, this is the kind of framework that makes the loop self-sustaining.
← Back to all articles
The framework runs a tight loop: a Researcher Agent proposes candidates using domain knowledge, an Engineer Agent builds and benchmarks them, an Analyzer Agent extracts reusable insights, and a Cognition Store accumulates all learned priors for the next iteration. Think of it as autoresearch with memory — every experiment makes the next one smarter.
The numbers are hard to argue with. In neural architecture search, ASI-Evolve discovered 105 SOTA linear attention architectures. The best one beat DeltaNet by 0.97 points — nearly 3x the gain of recent human-designed improvements. In data curation, the evolved pipeline improved benchmark performance by 3.96 points average, with MMLU gains exceeding 18 points. In RL algorithm design, discovered algorithms outperformed GRPO by up to 12.5 points on AMC32.
What makes this different from prior autoresearch is the Cognition Store. Instead of starting fresh each run, accumulated human priors and experimental insights carry forward. The system doesn't just search — it learns what kinds of architectures work and why, then uses that understanding to guide future exploration.
The code is fully open-source under Apache 2.0 at github.com/GAIR-NLP/ASI-Evolve. If the trend of AI designing AI holds, this is the kind of framework that makes the loop self-sustaining.
Comments