AI Designs AI: New Framework Automates Research

Researchers at SII-GAIR have introduced ASI-EVOLVE, a novel framework designed to automate the entire optimization cycle for AI research and development. This agentic system operates on a continuous loop, integrating prior knowledge, hypothesis generation, experimentation, and refinement. By automating the process of optimizing training data, model architectures, and learning algorithms, ASI-EVOLVE aims to overcome significant bottlenecks in AI innovation.

The framework comprises a 'Cognition Base' for domain expertise and an 'Analyzer' for complex feedback. Additional modules include a 'Researcher' for hypothesis generation, an 'Engineer' for running efficient experiments, and a 'Database' for persistent memory. This unified approach allows AI to learn from experimental feedback without constant human intervention.

Experiments demonstrated ASI-EVOLVE's capability in improving data curation, model architectures, and learning algorithms. For instance, AI-curated data led to significant score boosts on benchmark tests, and novel neural architecture designs surpassed human-designed baselines. The system also developed reinforcement learning algorithms that outperformed competitive baselines on complex mathematical reasoning tasks.

ASI-EVOLVE offers enterprises a way to integrate proprietary knowledge and iterate on internal AI systems, reducing the immense computational resources and engineering hours typically required for optimization. The researchers have open-sourced the ASI-EVOLVE code to make this foundational framework accessible.