This framework automates AI research by closing the loop between knowledge acquisition, hypothesis generation, experimentation, and analysis. It targets researchers and practitioners across various domains—from AI and biomedicine to finance and climate science—who can benefit from an autonomous agent that explores complex problem spaces to discover novel solutions more efficiently than manual methods.

How It Works

ASI-Evolve employs a four-step autonomous loop: LEARN (knowledge retrieval), DESIGN (candidate generation), EXPERIMENT (execution and metric collection), and ANALYZE (lesson distillation). Three core agents drive this process: the Researcher proposes candidates, the Engineer executes experiments, and the Analyzer synthesizes outcomes. Two memory systems are crucial: the Cognition Store injects domain knowledge and heuristics, while the Experiment Database logs all trials, enabling informed sampling strategies (UCB1, greedy, random, MAP-Elites) to prevent cycles and guide future exploration.

Quick Start & Requirements

• Primary install: pip install -r requirements.txt
• Prerequisites: Python 3.10+, bash and python3 on system path, any OpenAI-compatible API endpoint (e.g., GPT-4o, Claude, Gemini, local models via LiteLLM). Optional: Weights & Biases for experiment tracking.
• Demo: Includes a runnable circle-packing demo.
• Links: Paper (https://github.com/GAIR-NLP/ASI-Evolve/blob/main/assets/paper.pdf), arXiv (https://arxiv.org/abs/2603.29640).

Highlighted Details

• Achieved state-of-the-art results across challenging AI domains: 105 SOTA linear attention architectures (+0.97 pts), improved pretraining data curation (+3.96 pts avg, +18 pts MMLU), novel RL algorithm mechanisms (+12.5 pts AMC32), and stronger drug-target interaction models (+6.94 AUROC).
• Demonstrates a significant acceleration over manual research, generating 50-200 candidates per run compared to 5-10 ideas per week, with all insights systematically captured in a shared database and cognition store.
• The framework autonomously discovers novel optimization mechanisms and architectures, producing frontier-level results without human intervention in the loop.
• Reaches SOTA-level circle-packing results in as few as 17 rounds, showcasing rapid convergence on complex optimization tasks.

Maintenance & Community

• Community engagement is facilitated through GitHub Issues (https://github.com/GAIR-NLP/ASI-Evolve/issues) and Discussions (https://github.com/GAIR-NLP/ASI-Evolve/discussions). No specific details on core maintainers, sponsorships, or partnerships are provided in the README.

Licensing & Compatibility

The repository is presented as open-source, encouraging forking. However, a specific software license (e.g., MIT, Apache 2.0, GPL) is not explicitly stated in the README, which may require clarification for commercial use or integration into closed-source projects.

Limitations & Caveats

The README does not detail specific limitations. Potential adoption blockers could include the computational cost of extensive experimentation, reliance on the quality of underlying LLM APIs, and the necessity for a robust, well-defined evaluation script for the target problem.