Towards Optimal Circuit Generation: Multi-Agent Collaboration Meets Collective Intelligence

TL;DR

This work tackles the gate-level circuit design efficiency gap observed when using large language models by introducing CircuitMind, a six-agent, collaborative framework that distributes specialized reasoning across syntax-constrained generation, knowledge reuse, and dual optimization targets. Paired with TC-Bench, a gate-level benchmark built on collective intelligence from TuringComplete, the approach reframes hardware design as a multi-agent collaboration problem rather than a scaling problem. Empirical results show significant gains in both functional correctness and physical efficiency (SEI) across diverse models, with 55.6% of CircuitMind solutions matching or surpassing top human experts, including Phi-4 achieving top-tier performance despite smaller size. The study demonstrates that architectural innovations and human-aligned benchmarks, rather than merely larger models or more data, are key to achieving human-competitive circuit optimization, and provides an open-source platform for continued development in this direction.

Abstract

Large language models (LLMs) have transformed code generation, yet their application in hardware design produces gate counts 38\%--1075\% higher than human designs. We present CircuitMind, a multi-agent framework that achieves human-competitive efficiency through three key innovations: syntax locking (constraining generation to basic logic gates), retrieval-augmented generation (enabling knowledge-driven design), and dual-reward optimization (balancing correctness with efficiency). To evaluate our approach, we introduce TC-Bench, the first gate-level benchmark harnessing collective intelligence from the TuringComplete ecosystem -- a competitive circuit design platform with hundreds of thousands of players. Experiments show CircuitMind enables 55.6\% of model implementations to match or exceed top-tier human experts in composite efficiency metrics. Most remarkably, our framework elevates the 14B Phi-4 model to outperform both GPT-4o mini and Gemini 2.0 Flash, achieving efficiency comparable to the top 25\% of human experts without requiring specialized training. These innovations establish a new paradigm for hardware optimization where collaborative AI systems leverage collective human expertise to achieve optimal circuit designs. Our model, data, and code are open-source at https://github.com/BUAA-CLab/CircuitMind.

Figures (7)

• Figure 1: LLM-based EDA Evolution: From Behavioral to Physical Design
• Figure 2: TC-Bench's collective intelligence harvesting process
• Figure 3: CircuitMind System Architecture
• Figure 4: CoderAgent and Reviewer dynamic prompting process
• Figure 5: Overall Pass@1 and SEI Comparison: Base LLMs vs. CircuitMind.