AgentDSE: Reasoning-Augmented Architectural Design Space Exploration

AgentDSE introduces an LLM-driven methodology that automates architectural design space exploration by reasoning through physical constraints and performance dynamics, achieving competitive results with up to 100x fewer simulator evaluations than traditional methods. The approach eliminates the need for fine-tuning, precomputed databases, or domain-specific optimizers while producing interpretable decision traces.

AgentDSE represents a significant methodological shift in how complex system optimization can be approached through general-purpose AI reasoning rather than specialized algorithms. The core innovation lies in replacing black-box optimization with an LLM agent that mimics human architectural thinking—understanding physical constraints, identifying performance bottlenecks, analyzing data reuse patterns, and reasoning about workload structures. This mirrors broader trends in AI where foundation models are being applied to domain-specific problems without extensive customization.

The research demonstrates substantial efficiency gains across three distinct problem domains: DNN accelerator mapping, hardware/software co-design, and CPU cache optimization. Achieving comparable or superior design quality with orders of magnitude fewer evaluations addresses a genuine pain point in systems architecture. Traditional design space exploration consumes enormous computational resources, making the development cycle lengthy and expensive. By embedding reasoning capabilities into the search process, AgentDSE fundamentally changes the economics of architectural design.

For hardware and AI acceleration industries, this approach has immediate implications. Architects could dramatically reduce time-to-optimization for new hardware designs, accelerating product development cycles. The interpretable traces produced by AgentDSE also provide value beyond the final design—surfacing implicit assumptions, performance cliffs, and potential simulator artifacts that human designers can learn from. This transparency contrasts with black-box optimization methods where decision rationale remains opaque.

The methodology's applicability across diverse optimization domains suggests broader potential for LLM-based reasoning in systems engineering. Future work likely explores application to network architecture, compiler optimization, and other complex design spaces where human expertise has traditionally guided exploration.

• →AgentDSE achieves up to 100x reduction in simulator evaluations while maintaining or improving design quality across multiple architectural domains.
• →The approach requires no model fine-tuning, precomputed design databases, or custom optimizer code, making it broadly applicable.
• →Interpretable decision traces from the LLM agent expose architectural hypotheses and performance dynamics rather than burying logic in optimizer state.
• →The methodology demonstrates that general-purpose LLMs can effectively replicate human architectural reasoning for complex design problems.
• →The approach potentially accelerates hardware development cycles and reduces computational costs associated with traditional design space exploration.