New AI framework autonomously optimizes training data, architectures and algorithms — outperforming human baselines
Publish Date: 2026-04-27 11:56:00
Source Domain: venturebeat.com
- A new framework called ASI-EVOLVE, developed by researchers at the Generative Artificial Intelligence Research Lab (SII-GAIR), aims to automate the optimization of training data, model architectures, and learning algorithms to close the bottleneck in AI research and development (R&D).
- ASI-EVOLVE operates through a continuous cycle of learning, designing hypotheses, experimenting, and analyzing outcomes, which it continuously feeds back into its knowledge base for growth and improvement.
- The framework autonomously discovered novel designs that outperformed human-designed baselines, improving pretraining data pipelines, generating new language model architectures, and designing efficient reinforcement learning algorithms.
- Components of ASI-EVOLVE include the Cognition Base (foundational expertise storage), the Analyzer (interpreting experimental feedback), the Researcher agent (hypothesis generation), the Engineer (running experiments), and the Database (persistent memory).
- ASI-EVOLVE effectively improves data curation, model architecture design, and learning algorithm design, enhancing the performance of AI models as demonstrated in benchmark tests.
- Enterprises benefit from the framework by reducing manual engineering effort while attaining performance levels comparable or superior to human-designed models, potentially integrating proprietary knowledge to further optimize internal AI systems.
- The research team has open-sourced ASI-EVOLVE, making the framework accessible to developers and product builders.
