Agentic evolution of physically constrained foundation models
06:00 · June 25, 2026 · arXiv cs.AI RSS
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.
Summary
A multi-agent discovery engine addresses a persistent limitation in automated AI research: generalist agents frequently generate hardware-incompatible designs because they lack explicit physical grounding. The system counters this by anchoring its reasoning in an Evolutionary Knowledge Graph built from 164 existing large-language-model compression techniques. This graph records both macro-level dependencies between methods and micro-level evolutionary checkpoints, allowing the engine to extract an algorithmic Chain-of-Thought that converts prior successful optimizations into directed structural mutations rather than relying on blind stochastic search.
Applied to the demanding task of compressing foundation models under strict memory and bandwidth constraints, the engine produced two hardware-aware techniques. Q-Enhance dynamically reallocates numerical precision to mitigate accuracy loss during long-context inference in dense models. MoE-Salient-AQ introduces an expert-granularity quantization scheme combined with adaptive low-rank compensation, improving upon prior state-of-the-art sparse Mixture-of-Experts designs by 3.7 percent in sub-3-bit regimes. A bandwidth-efficient Sensitivity Profile supplies direct hardware feedback that steers the evolutionary process toward feasible deployment configurations.
The resulting compression pipeline enabled a 235-billion-parameter model to run on a dual-A100 server, cutting memory requirements from 438 GB to 108 GB—a 75 percent reduction—while incurring only a 0.64 percent accuracy drop. The same framework supports smaller models on single consumer GPUs, demonstrating that the approach scales across edge and cloud regimes. By replacing unconstrained combinatorial exploration with knowledge-guided autonomy, the work establishes a repeatable hardware–software co-design loop for machine-driven discovery within concrete physical limits.
Why it matters
This research is highly relevant for Dutch AI researchers and infrastructure engineers focusing on efficient, sustainable AI deployment. By drastically reducing the hardware requirements for massive foundation models, it enables local, cost-effective deployment for SMEs and aligns with European goals for green AI and data sovereignty.
