Meta Paper Uses Two-Agent Search to Find Architectures That Beat Llama 3.2 in 24 Hours
Meta published a paper titled "Agentic Discovery of Neural Architectures" describing AIRA, a two-agent system that autonomously searches for neural architectures that outperform Llama 3.2 at 350M, 1B and 3B scales within a 24-hour compute budget.
The framework splits the work between AIRA-Compose, which searches the macro architecture, and AIRA-Design, which implements lower-level mechanisms. The paper said this approach outperformed a single end-to-end agent on the search task and could also be applied to pipeline assembly, query planning, prompt scaffolding and tool-use programs.
From the sources (2 posts)
@dair_aiNEW paper from Meta: Agentic Discovery of Neural Architectures. This is a hot new area of research! Keep an eye on it.
@omarsar0NEW paper from Meta. (bookmark it) It's an agent system that autonomously discovers neural architectures that beat Llama 3.2 at 350M, 1B, and 3B scales, all under a 24-hour compute budget. They get this work by splitting the search into