Order Matters: Unveiling the Hidden Impact of Macro Placement Sequences via Proxy-Guided LLM Evolution

A new framework called OrderPlace uses large language models to automatically discover the sequence in which macro blocks are placed during chip design, a step researchers say has been overlooked despite its critical impact on final quality [1]. Macro placement is a fundamental stage in modern chip physical design, determining how large functional blocks are arranged on a silicon die [1]. While machine learning has increasingly been applied to decide where these blocks go, the order in which they are placed has remained governed by static, hand-crafted heuristics [1]. The authors of the new paper argue that this temporal dimension is not a minor preprocessing detail but a decisive factor, where poor early choices create irreversible domino effects that shrink the available solution space [1]. To address this, the researchers propose OrderPlace, a proxy-guided LLM evolution framework that explores a broader space of code-level policies for placement sequencing [1]. Instead of relying on conventional rules such as ordering by area or connectivity, OrderPlace can generate strategies ranging from static scoring metrics to dynamic, physics-inspired mechanisms [1]. Because exhaustively evaluating every candidate sequence is computationally prohibitive, the framework uses a lightweight proxy mechanism that filters candidates with a deterministic greedy probe before committing to full evaluation [1]. The system was tested on the standard ISPD 2005 benchmarks, a widely used set of circuits for evaluating physical design tools [1]. OrderPlace reduced total wirelength by 34.04% compared with WireMask-EA and by 14.08% compared with EGPlace, the prior state-of-the-art method [1]. The work was submitted to arXiv on June 8, 2026 [2]. The use of large language models for code generation and optimization has expanded rapidly across scientific domains. A recent scoping review of generative systems for quantum computing, for instance, catalogued thirteen systems that produce Qiskit code, OpenQASM programs, or circuit graphs, though none had been evaluated end-to-end on real quantum hardware [3]. The broader LLM landscape includes models such as DeepSeek, a Chinese system trained at a reported cost of US$6 million, far below the sums spent by competitors [7], and Alibaba Cloud’s Qwen family, distributed under open-source licenses including Apache 2.0 [9]. The OrderPlace paper appears on arXiv, a preprint server that has integrated with platforms such as Hugging Face to link papers directly to interactive demos and community-built Spaces [5][6]. This integration allows readers to try models without writing code and has been available since November 2022 [5].