Agentic evolution: physically constrained foundation models

Agentic evolution of physically constrained foundation models introduces a physically grounded, multi-agent discovery engine that autonomously architects hardware-compliant computing systems. The paper, authored by Jiangwei Zhang and nine coauthors and submitted on 24 Jun 2026, centers an Evolutionary Knowledge Graph and an "algorithmic Chain-of-Thought" to guide search and produce hardware-aware compression methods.

What did the paper do?

The paper describes a multi-agent discovery engine that converts blind stochastic search into directed structural evolution using an Evolutionary Knowledge Graph and an "algorithmic Chain-of-Thought", and it reports concrete results on compression and deployment. The engine produced two hardware-aware compression methods, Q-Enhance and MoE-Salient-AQ, and the authors applied them to foundation-model deployment and constrained hardware.

The submission runs 29 pages and includes 5 main figures plus 4 extended data figures. The authors report that MoE-Salient-AQ outperforms state-of-the-art manual sparse Mixture-of-Experts designs by 3.7% in sub-3-bit regimes, and that Q-Enhance mitigates long-context accuracy loss in dense models. They also present a bandwidth-efficient Sensitivity Profile used to guide deployment.

How does the engine and the methods work?

The engine structures past innovations into an Evolutionary Knowledge Graph, extracts an "algorithmic Chain-of-Thought" to direct search, and uses a Sensitivity Profile to prioritize bandwidth and memory trade-offs during co-design. Agents autonomously propose and evaluate architectures, converting combinatorial exploration into knowledge-driven evolution.

From that setup the paper details two resulting methods. Q-Enhance addresses long-context accuracy degradation in dense models; MoE-Salient-AQ is a hardware-aware sparse Mixture-of-Experts variant optimized for low-bit regimes. The Sensitivity Profile guides which parameters to compress or sparsify to meet strict physical constraints, enabling the system to balance accuracy and resource budgets.

How well did the new methods perform and what deployment did they demonstrate?

MoE-Salient-AQ beat manual sparse MoE designs by 3.7% at sub-3-bit quantization, and the team deployed a massive 235-billion-parameter model onto a constrained dual-A100 server with a 75% reduction in memory requirement and a 0.64% accuracy degradation. Those are the headline empirical numbers provided by the authors.

The paper frames those results as examples of the engine converting unconstrained combinatorial search into directed, scalable hardware-software co-design, using sensitivity and knowledge graph signals to meet physical limits while preserving model performance.

Why it matters

Hardware constraints frequently block automated scientific discovery because models and proposed architectures can be infeasible on real devices. By encoding past innovations into an Evolutionary Knowledge Graph and extracting a directed "algorithmic Chain-of-Thought", the paper demonstrates an automated path to hardware-compliant designs rather than human guesswork. The concrete numbers—3.7% improvement for MoE-Salient-AQ in sub-3-bit regimes and a 75% memory cut for a 235-billion-parameter model with only 0.64% accuracy loss—show the approach can produce practical, deployable gains under strict physical boundaries.

What to watch

Look for the authors to publish code, datasets, or reproducibility materials tied to the Sensitivity Profile and the two methods, and for follow-up experiments that validate MoE-Salient-AQ and Q-Enhance across other hardware configurations. The next confirmatory signal would be independent replication of the 235-billion-parameter deployment on constrained dual-A100 hardware and the reported 3.7% gain in sub-3-bit regimes.