GR-Evolve: Design-Adaptive Global Routing via LLM-Driven Algorithm Evolution Conference: ; ;
Taizun Jafri, Vidya A. Chhabria · 2026
Read GR-Evolve: Design-Adaptive Global Routing via LLM-Driven Algorithm Evolution Conference: ; ; by Taizun Jafri, Vidya A. Chhabria (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-22234.txt.
Introduction
Jafri and Chhabria introduce design-adaptive chip-design tooling, in which a tool's own algorithms are tuned to each individual circuit rather than left fixed. Their GR-Evolve system uses an AI agent to rewrite global routing code step by step, guided by quality feedback inside the OpenROAD toolchain. Across seven benchmark designs on three technology nodes, it cut post-routing wirelength by up to 8.72 percent over existing baseline routers.
Catalog record
| CREATOR | Taizun Jafri, Vidya A. Chhabria |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55693 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.22234
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GR-Evolve: Design-Adaptive Global Routing via LLM-Driven Algorithm Evolution Conference: ; ; Taizun Jafri, Vidya A. Chhabria Abstract . Modern ASIC design is becoming increasingly complex, driving up design costs while limiting productivity gains from existing EDA tools. Despite decades of progress, current tools rely on fixed heuristics and expose limited control through tool hyperparameters that require extensive manual tuning to achieve acceptable quality of results (QoR). While prior work has explored learning-based optimization and design-specific hyperparameter tuning, these approaches operate within the constraints of static tool algorithm implementations and do not adapt the underlying algorithms to individual designs. To address this limitation, we introduce the concept of design-adaptive EDA tooling, in which the internal algorithms of EDA tools are automatically specialized to the characteristics of a given design. We instantiate this paradigm through GR-Evolve, a code evolution framework that leverages an agentic large language model (LLM) to iteratively modify global routing source code using QoR-driven feedback. The framework equips the LLM with persistent contextual knowledge of open-source global routers along with an integrated toolchain for QoR evaluation within the OpenROAD infrastructure. We evaluate GR-Evolve across seven benchmark designs across three technology nodes and demonstrate up to 8.72% reduction in post-detailed-routing wirelength over existing baseline routers, highlighting the potential of LLM-driven EDA code evolution for design-adaptive global routing. 1. Introduction Motivation. Global routing is a critical stage in the physical design flow of integrated circuits, where approximate routing paths are determined for all nets across the chip. During this phase, the routing region is partitioned into a grid of global routing cells, called GCells, and each net is assigned a sequence of GCells through which it will pass, without specifying the exact locations of the wires. The primary objective is to efficiently allocate routing resources while minimizing congestion, wirelength, and delay. Global routing directly impacts the routability and performance of the final design. As modern chip designs continue to scale in complexity, effective global routing has become increasingly important to design closure. Figure 1 . Design-adaptive EDA tools via LLM-driven EDA tool source code generation. Prior global routing work. Modern academic global routers ( 19 ; 14 ; 11 ) employ diverse algorithmic frameworks to address large-scale routing challenges. FastRoute ( 19 ) employs an efficient rip-up-and-reroute scheme that combines congestion-driven Steiner tree construction with 3-bend routing to achieve fast convergence. CUGR ( 14 ) operates directly in 3D space, leveraging 3D pattern routing and multi-level maze routing guided by a probabilistic resource model to improve detailed routability. SPRoute ( 11 ) introduces a soft-capacity mechanism that reserves routing resources based on local pin density and employs parallelization for runtime efficiency. Despite these advances, all existing global routers rely heavily on human-designed heuristics and fixed cost functions to explore the routing solution space. As a result, their performance is inherently design-dependent, and no single router consistently achieves the best quality of results (QoR) and runtime across all benchmarks. This highlights a fundamental limitation of current EDA practice: the use of static tools that are applied uniformly across diverse designs (one-size fits all). While one could attempt to tune hyperparameters within a given tool, these approaches remain limited. They do not enable fine-grained adaptation of the underlying algorithms, nor do they allow discovery of new heuristics tailored to specific design characteristics. Consequently, a significant portion of achievable QoR remains unexplored. To the best of our knowledge no prior global routing work has explored LLM-driven source code evolution to automatically generate design-adaptive routing algorithms, leaving this opportunity largely untapped. In this work, we address this limitation by recasting global routing as a program optimization problem. Instead of optimizing routing solutions or tuning parameters of a fixed tool, we seek to optimize the routing algorithm itself. Specifically, given a baseline global router and a target design, the goal is to automatically evolve the router’s source code to produce a design-specialized algorithm that improves QoR under full routing flows. This formulation shifts the focus from design space exploration to design–tool co-exploration , where both the design and the underlying EDA tool are jointly optimized. Opportunity. Large language models (LLMs) have undergone rapid evolution in the broader artificial intelligence (AI) community. LLMs have been integrated into agentic systems that interact directly with external tools ( 4 ; 2 ) , execute actions, observe outcomes, and iteratively refine behavior through feedback. This idea was advanced by AlphaEvolve ( 15 ) , which integrates LLMs into large-scale evolutionary optimization pipelines that operate over entire codebases and algorithmic systems. By coupling LLM-driven proposal generation with automated evaluation, selection, and iteration, AlphaEvolve marks a transition from LLMs as interactive assistants to LLMs as autonomous components. These advances have been adopted in some early works for physical design and SAT. AuDoPEDA ( 8 ) formalizes a closed-loop autonomous coding pipeline that systematically improves QoR for the open-source OpenROAD EDA stack by combining documentation, planning, and execution. However, this work focuses on placement and timing optimization and does not modify the global router. Another recent work ( 27 ) , inspired by AlphaEvolve, has also demonstrated the feasibility of extending LLM-driven code evolution to full repository scale. SATLUTION ( 27 ) is an example of such a framework, applying LLM-based evolutionary optimization to Boolean satisfiability. Our work. In this work, we recast global routing as a program optimization problem, where the objective is to automatically evolve the source code of a routing algorithm to specialize it for a target design. We introduce the concept of design-adaptive EDA tools , enabled by LLMs that generate and evolve repository-scale global routing code to improve QoR through design–tool co-exploration . Fig. 1 illustrates this paradigm. We instantiate this concept in the context of global routing. We present GR-Evolve, an agentic framework that leverages an LLM-powered command-line interface ( 4 ) to iteratively evolve global router source code, automatically specializing routing heuristics and cost functions within the router implementation to the characteristics of a given design. The framework provides the agent with persistent contextual knowledge, including three open-source global router codebases, concise summaries of their routing strategies, and an integrated toolchain for building ( 1 ) and executing the OpenROAD-flow-scripts (ORFS) on placed designs. In a closed-loop cycle, the agent reasons about routing heuristics, modifies source code, compiles the updated router within OpenROAD, and evaluates its performance on a target design. Through successive iterations, the framework synthesizes specialized global routers tailored to patterns of individual designs. GR-Evolve performs a bounded exploration of the program space and selects the best-performing router observed during the search, rather than attempting to converge to a single optimal solution. We validate this approach on seven benchmark designs, ranging from 15k to 270k nets, across three technology nodes, generating variants from three baseline routers. Our contributions are: • We are the first to formulate global routing as a program-level optimization problem over source code. • We propose a design-adaptive agentic framework that automatically evolves global routing algorithms on a per-design (design-tool co-exploration) basis using LLM-driven code modification. • We develop a stateless, version-controlled evolution framework that externalizes all state into persistent artifacts, enabling scalable long-horizon optimization. • We formulate router evolution as a multi-objective optimization problem that reduces detailed routing (DR) wirelength