{
  "@context": {
    "@vocab": "http://schema.org/",
    "hasPart": {
      "@id": "http://schema.org/hasPart",
      "@type": "@id"
    },
    "mainEntity": {
      "@id": "http://schema.org/mainEntity",
      "@type": "@id"
    },
    "about": {
      "@id": "http://schema.org/about",
      "@type": "@id"
    },
    "citation": {
      "@id": "http://schema.org/citation",
      "@type": "@id"
    },
    "author": {
      "@id": "http://schema.org/author",
      "@type": "@id"
    },
    "publisher": {
      "@id": "http://schema.org/publisher",
      "@type": "@id"
    },
    "creator": {
      "@id": "http://schema.org/creator",
      "@type": "@id"
    },
    "contributor": {
      "@id": "http://schema.org/contributor",
      "@type": "@id"
    },
    "Question": "http://schema.org/Question",
    "Answer": "http://schema.org/Answer",
    "DefinedTerm": "http://schema.org/DefinedTerm",
    "DefinedTermSet": "http://schema.org/DefinedTermSet",
    "HowTo": "http://schema.org/HowTo",
    "HowToStep": "http://schema.org/HowToStep"
  },
  "@type": "ScholarlyArticle",
  "@id": "#ASI-Evolve",
  "headline": "ASI-Evolve: AI Accelerates AI",
  "author": [
    {
      "@type": "Person",
      "name": "Weixian Xu",
      "affiliation": [
        {"@id": "#SJTU"},
        {"@id": "#SII"},
        {"@id": "#GAIR"}
      ],
      "role": "Leading author"
    },
    {
      "@type": "Person",
      "name": "Tiantian Mi",
      "affiliation": [
        {"@id": "#SJTU"},
        {"@id": "#SII"},
        {"@id": "#GAIR"}
      ],
      "role": "Core contributor"
    },
    {
      "@type": "Person",
      "name": "Yixiu Liu",
      "affiliation": [
        {"@id": "#SJTU"},
        {"@id": "#SII"},
        {"@id": "#GAIR"}
      ],
      "role": "Core contributor"
    },
    {
      "@type": "Person",
      "name": "Yang Nan",
      "affiliation": [
        {"@id": "#SII"},
        {"@id": "#GAIR"}
      ],
      "role": "Core contributor"
    },
    {
      "@type": "Person",
      "name": "Zhimeng Zhou",
      "affiliation": [
        {"@id": "#SII"}
      ]
    },
    {
      "@type": "Person",
      "name": "Lyumanshan Ye",
      "affiliation": [
        {"@id": "#SJTU"},
        {"@id": "#SII"},
        {"@id": "#GAIR"}
      ]
    },
    {
      "@type": "Person",
      "name": "Lin Zhang",
      "affiliation": [
        {"@id": "#SJTU"},
        {"@id": "#SII"},
        {"@id": "#GAIR"}
      ]
    },
    {
      "@type": "Person",
      "name": "Yu Qiao",
      "affiliation": [
        {"@id": "#SII"}
      ]
    },
    {
      "@type": "Person",
      "name": "Pengfei Liu",
      "affiliation": [
        {"@id": "#SJTU"},
        {"@id": "#SII"},
        {"@id": "#GAIR"}
      ],
      "role": "Corresponding author"
    }
  ],
  "publisher": {
    "@type": "Organization",
    "name": "SII-GAIR",
    "url": "https://github.com/GAIR-NLP/ASI-Evolve"
  },
  "datePublished": "2026-03-31",
  "url": "https://arxiv.org/abs/2603.29640v1",
  "abstract": "ASI-EVOLVE is an agentic framework for AI-for-AI research that closes the loop of learn–design–experiment–analyze to accelerate AI development. It integrates a cognition base injecting human priors and an analyzer distilling experimental outcomes into reusable insights. Demonstrated across three core AI development components—data, architectures, and learning algorithms—ASI-EVOLVE discovers 105 SOTA linear attention architectures surpassing human designs, improves pretraining data curation with +3.96 average benchmark gains, and designs reinforcement learning algorithms outperforming GRPO by up to +12.5 points. It also shows transferability beyond AI in mathematics and biomedicine, suggesting feasibility of closed-loop AI research.",
  "articleBody": "ASI-EVOLVE is an agentic framework for AI-for-AI research that closes the loop of learn–design–experiment–analyze to accelerate AI development. It integrates a cognition base injecting human priors and an analyzer distilling experimental outcomes into reusable insights. Demonstrated across three core AI development components—data, architectures, and learning algorithms—ASI-EVOLVE discovers 105 SOTA linear attention architectures surpassing human designs, improves pretraining data curation with +3.96 average benchmark gains, and designs reinforcement learning algorithms outperforming GRPO by up to +12.5 points. It also shows transferability beyond AI in mathematics and biomedicine, suggesting feasibility of closed-loop AI research.",
  "hasPart": [
    {
      "@type": "DefinedTermSet",
      "@id": "#DefinedTerms",
      "name": "Defined Terms in ASI-EVOLVE",
      "description": "Key terms defined in the context of ASI-EVOLVE framework and AI research.",
      "hasDefinedTerm": [
        {
          "@type": "DefinedTerm",
          "name": "ASI-EVOLVE",
          "description": "An agentic framework for AI-for-AI research that closes the loop of learn–design–experiment–analyze to accelerate AI development."
        },
        {
          "@type": "DefinedTerm",
          "name": "Cognition Base",
          "description": "A knowledge repository encoding human prior knowledge, heuristics, and design principles to guide AI exploration."
        },
        {
          "@type": "DefinedTerm",
          "name": "Analyzer",
          "description": "A module that distills complex experimental outcomes into compact, actionable insights for future iterations."
        },
        {
          "@type": "DefinedTerm",
          "name": "Learn–Design–Experiment–Analyze Cycle",
          "description": "The iterative scientific process cycle implemented by ASI-EVOLVE for autonomous AI research."
        },
        {
          "@type": "DefinedTerm",
          "name": "Model Architecture Design",
          "description": "The task of designing neural network architectures, specifically efficient linear attention mechanisms."
        },
        {
          "@type": "DefinedTerm",
          "name": "Pretraining Data Curation",
          "description": "The task of designing data cleaning and curation strategies to improve pretraining dataset quality."
        },
        {
          "@type": "DefinedTerm",
          "name": "Reinforcement Learning Algorithm Design",
          "description": "The task of designing novel RL algorithms for training large language models."
        },
        {
          "@type": "DefinedTerm",
          "name": "Scientific Task Length (Ltask)",
          "description": "An analytical framework characterizing autonomous scientific research tasks by execution cost, search space complexity, and feedback complexity."
        },
        {
          "@type": "DefinedTerm",
          "name": "UCB1 Sampling",
          "description": "A database sampling algorithm balancing exploration and exploitation using upper confidence bounds."
        },
        {
          "@type": "DefinedTerm",
          "name": "MAP-Elites Sampling",
          "description": "A sampling algorithm maintaining a quality-diversity archive to preserve diverse solution niches."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "#HowToUseASI-EVOLVE",
      "name": "How to Use ASI-EVOLVE Framework",
      "description": "Steps to perform autonomous AI research using ASI-EVOLVE.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Initialize Cognition Base",
          "text": "Collect and encode human prior knowledge, heuristics, and relevant literature into the cognition repository."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Generate Candidate Programs",
          "text": "Sample context nodes and retrieve cognition items to condition the Researcher module to generate new candidate AI designs or algorithms."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Execute Experiments",
          "text": "Run the candidate programs in the experimental environment to obtain evaluation metrics and scalar fitness scores."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Analyze Experimental Outcomes",
          "text": "Use the Analyzer to distill complex experimental logs and metrics into concise, actionable reports."
        },
        {
          "@type": "HowToStep",
          "position": 5,
          "name": "Update Database and Iterate",
          "text": "Store motivations, programs, results, and analyses in the database; use these to inform subsequent rounds of exploration."
        }
      ]
    },
    {
      "@type": "Question",
      "@id": "#Q1",
      "name": "Can AI accelerate the development of AI itself?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ASI-EVOLVE demonstrates that AI can accelerate AI development by autonomously discovering improvements in data, architectures, and learning algorithms through a closed-loop learn–design–experiment–analyze cycle."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q2",
      "name": "What are the three central components of AI development targeted by ASI-EVOLVE?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Data curation, model architecture design, and reinforcement learning algorithm design."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q3",
      "name": "How does ASI-EVOLVE incorporate human knowledge?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Through a cognition base that encodes human prior knowledge, heuristics, and design principles, which are retrieved and injected into each round of exploration."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q4",
      "name": "What role does the Analyzer play in ASI-EVOLVE?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It distills complex, multi-dimensional experimental outcomes into compact, actionable insights that guide future iterations."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q5",
      "name": "What are the benefits of using ASI-EVOLVE for model architecture design?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ASI-EVOLVE discovered 105 SOTA linear attention architectures surpassing human designs, with top models achieving nearly triple the improvement of recent human-designed architectures."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q6",
      "name": "How does ASI-EVOLVE improve pretraining data curation?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "By designing category-specific cleaning strategies that improve data quality, resulting in average benchmark improvements of +3.96 points and gains exceeding 18 points on knowledge-intensive tasks like MMLU."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q7",
      "name": "What improvements does ASI-EVOLVE achieve in reinforcement learning algorithm design?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It discovers novel RL algorithms that outperform the GRPO baseline by up to +12.5 points on AMC32, +11.67 points on AIME24, and +5.04 points on OlympiadBench."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q8",
      "name": "Can ASI-EVOLVE's AI-for-AI paradigm transfer beyond AI development?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Yes, initial evidence shows transferability to broader scientific domains such as mathematics and biomedicine, including drug-target interaction prediction."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q9",
      "name": "What is the Scientific Task Length (Ltask) framework?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "An analytical framework characterizing autonomous scientific research tasks by execution cost, search space complexity, and feedback complexity."
      }
    },
    {
      "@type": "Question",
      "@id": "#Q10",
      "name": "How does ASI-EVOLVE compare to other evolutionary frameworks on the circle packing task?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "ASI-EVOLVE reaches state-of-the-art performance faster and with fewer evolution steps than frameworks like AlphaEvolve, OpenEvolve, and GEPA."
      }
    }
  ],
  "hasPart": [
    {
      "@type": "HowTo",
      "@id": "#HowToImproveAIArchitecture",
      "name": "How to Improve AI Model Architecture Using ASI-EVOLVE",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Initialize Cognition Repository",
          "text": "Load domain priors from literature on linear attention, state space models, and efficient transformers."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Generate Candidate Architectures",
          "text": "Sample top-performing nodes and retrieve cognition context to generate new candidate attention architectures."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Evaluate Candidates",
          "text": "Train small models on benchmarks, score candidates combining quantitative metrics and LLM qualitative scores."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Verify and Scale Promising Models",
          "text": "Scale promising architectures to larger models and validate gains on extended benchmarks."
        },
        {
          "@type": "HowToStep",
          "position": 5,
          "name": "Analyze and Iterate",
          "text": "Analyze top architectures for adaptive routing mechanisms and iterate to discover improved designs."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "#HowToCuratePretrainingData",
      "name": "How to Curate Pretraining Data Using ASI-EVOLVE",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Initialize Cognition Repository",
          "text": "Identify data quality issues such as HTML artifacts, incomplete fragments, and domain-specific noise."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Generate Curation Strategies",
          "text": "Generate candidate data cleaning strategies addressing identified issues."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Execute Cleaning and Evaluate",
          "text": "Apply strategies on sampled documents and evaluate cleaned data quality with diagnostic feedback."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Refine Strategies",
          "text": "Incorporate diagnostic feedback to iteratively refine and improve curation strategies."
        },
        {
          "@type": "HowToStep",
          "position": 5,
          "name": "Train Models and Validate",
          "text": "Train models on curated data and validate improvements on benchmark tasks."
        }
      ]
    },
    {
      "@type": "HowTo",
      "@id": "#HowToDesignRLAlgorithm",
      "name": "How to Design Reinforcement Learning Algorithms Using ASI-EVOLVE",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Initialize Cognition Repository",
          "text": "Load recent research papers on variance reduction and KL-penalty modifications."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Generate Candidate Algorithms",
          "text": "Generate candidate RL algorithms modifying advantage allocation and gradient computation."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Evaluate Candidates",
          "text": "Train candidates on mathematical benchmarks and score based on accuracy and qualitative coherence."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Scale and Verify",
          "text": "Scale promising algorithms to larger models and verify improvements on extended evaluation suites."
        },
        {
          "@type": "HowToStep",
          "position": 5,
          "name": "Analyze Innovations",
          "text": "Analyze top algorithms for theoretical innovations such as pairwise asymmetric optimization and budget-constrained updates."
        }
      ]
    }
  ],
  "citation": [
    {
      "@type": "ScholarlyArticle",
      "name": "AlphaEvolve: A coding agent for scientific and algorithmic discovery",
      "author": "Alexander Novikov et al.",
      "datePublished": "2025",
      "url": "https://arxiv.org/abs/2408.06292"
    },
    {
      "@type": "ScholarlyArticle",
      "name": "Nemotron-CC-Math: A 133 billion-token-scale high quality math pretraining dataset",
      "author": "Rabeeh Karimi Mahabadi et al.",
      "datePublished": "2025"
    },
    {
      "@type": "ScholarlyArticle",
      "name": "Scaling deep learning for materials discovery",
      "author": "Amil Merchant et al.",
      "datePublished": "2023",
      "journalName": "Nature",
      "volumeNumber": "624",
      "pages": "80-85"
    },
    {
      "@type": "ScholarlyArticle",
      "name": "Highly accurate protein structure prediction with AlphaFold",
      "author": "John Jumper et al.",
      "datePublished": "2021",
      "journalName": "Nature",
      "volumeNumber": "596",
      "pages": "583-589"
    }
  ],
  "mainEntity": "#ASI-Evolve"
}