Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design
The paper introduces AIRA-Compose and AIRA-Design, two frameworks for the autonomous design of neural architectures. These frameworks enable AI agents to discover and optimize models that outperform traditional designs. This research marks a significant advancement towards recursive self-improvement in AI systems.
- ▪AIRA-Compose and AIRA-Design are dual-frameworks for designing neural architectures autonomously.
- ▪The frameworks utilize multiple agents to explore and evaluate millions of model candidates, leading to the discovery of 14 new architectures.
- ▪AIRAformers and AIRAhybrids consistently outperform existing models like Llama 3.2 in various tasks.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.15871 (cs) [Submitted on 15 May 2026] Title:Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design Authors:Alberto Pepe, Chien-Yu Lin, Despoina Magka, Bilge Acun, Yannan Nellie Wu, Anton Protopopov, Carole-Jean Wu, Yoram Bachrach View a PDF of the paper titled Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design, by Alberto Pepe and 7 other authors View PDF HTML (experimental) Abstract:Toward recursive self-improvement, we investigate LLM agents autonomously designing foundation models beyond standard Transformers. We introduce a dual-framework approach: AIRA-Compose for high-level architecture search, and AIRA-Design for low-level mechanistic implementation.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
