Agentic evolution of physically constrained foundation models

The problem is that contemporary generalist AI agents lack physical grounding, leading to hallucinated hardware-incompatible designs. The method introduces a physically grounded, multi-agent discovery engine that uses an Evolutionary Knowledge Graph and algorithmic Chain-of-Thought to direct structural evolution. Experimental evidence shows the engine evolved two compression methods—Q-Enhance and MoE-Salient-AQ—that surpass human heuristics, and deployed a 235-billion-parameter model on a dual-A100 server with 75% memory reduction and only 0.64% accuracy loss. This matters because it establishes a scalable hardware-software co-design paradigm for machine-driven discovery within strict physical constraints.