Agentic evolution of physically constrained foundation models

2026-06-24

AI Summary

This work addresses the reliability limitations of general-purpose AI agents in automated scientific discovery, which often generate hardware-incompatible designs due to a lack of physical constraints. To overcome this, the authors propose the first physics-aware multi-agent discovery engine that leverages an evolutionary knowledge graph to extract algorithmic chains of thought, transforming random search into directed structural evolution for autonomously designing hardware-compatible computational systems. The approach introduces a novel knowledge-driven structural evolution mechanism and presents two model compression techniques—Q-Enhance and MoE-Salient-AQ—that surpass the limitations of human-designed heuristics. Validated on a dual-A100 server, the framework successfully deploys a 235-billion-parameter model with a 75% reduction in memory footprint and only a 0.64% accuracy drop; notably, MoE-Salient-AQ achieves a 3.7% improvement over the best handcrafted sparse design at sub-3-bit precision.

Abstract

Artificial intelligence increasingly drives automated scientific discovery, yet contemporary generalist agents lack physical grounding, frequently hallucinating hardware-incompatible designs. Here, we present a physically grounded, multi-agent discovery engine that autonomously architects hardware-compliant computing systems. Anchored by an Evolutionary Knowledge Graph structuring past scientific innovations, the framework extracts an "algorithmic Chain-of-Thought" to transform blind stochastic search into directed structural evolution. Applied to the extreme testbed of foundation model deployment, the engine evolved two hardware-aware compression methodologies surpassing human-engineered heuristics: Q-Enhance mitigates long-context accuracy loss in dense models, and MoE-Salient-AQ outperforms state-of-the-art manual sparse Mixture-of-Experts designs by 3.7% at sub-3-bit regimes. Utilizing a bandwidth-efficient Sensitivity Profile, we successfully deployed a massive 235-billion-parameter model onto a constrained dual-A100 server, reducing memory requirements by 75% with a marginal 0.64% accuracy degradation. By transforming unconstrained combinatorial search into knowledge-driven autonomy, this establishes a scalable hardware-software co-design paradigm for machine-driven discovery within strict physical boundaries.

Problem

• physical grounding
• hardware compatibility
• foundation model deployment
• resource-constrained systems
• automated scientific discovery

Innovation

• physically grounded AI
• evolutionary knowledge graph
• hardware-aware compression
• algorithmic chain-of-thought
• hardware-software co-design