PCB routing reasoning benchmark

OmniRouting: A Semantic-Coupled Multimodal Benchmark for Constraint-Aware Spatial Reasoning in PCB Routing

Taiting Lu1*, Kaiyuan Lin1*, Ziwei Dong2, Sisong Bei2, Haolin Ye1, Yuxin Tian2, Runze Liu1, Mingjia Wang3, Jingying Zeng2, Hongxing Pan3, Kai Zhang3, Haoyu Wang1, Guoliang Shi3, Ling Ma3, Yifan Yang4, Jiaying Lu5, Qi He2, Yi-Chao Chen3, Sung-Liang Chen3, Yincheng Jin6, Mahanth Gowda†1

1Pennsylvania State University  |  2Independent Researcher  |  3Shanghai Jiao Tong University  |  4Microsoft Research  |  5Emory University  |  6Binghamton University

1,681 industrial-grade, schematic-coupled PCB designs for evaluating connectivity, design-rule compliance, electrical functionality, and tool-augmented agentic routing.

Paper PDF

Agentic evaluation

Qualitative Routing Trajectories

Human-engineered reference routing and model-generated trajectories for the Adafruit Bluefruit LE UART Friend under the No Tools setting, shown side by side for comparison of path planning, connectivity, and design-rule compliance.

Board outline Through-hole pads Top-side pads / components Back-side pads / components Routed traces

Paper results

OmniRouting leaderboards

Complete one-shot and agentic routing results from the current manuscript. Unavailable results are preserved as dashes, matching the paper.

NRR net routability rate (%) PRR pad-pair routability rate (%) PR pass rate Open / PShort / NShort affected nets (%) Total / Clr. / Dist. mean violations Polygon outputs with a copper pour (%) TWL routed-wire centerline length RT runtime (s)

Table 1 - One-shot Evaluation of LMMs on PCB Routing

PR is the fraction of expected inputs producing a loadable output. NRR is the percentage of required nets that are connected and free of published design-rule violations; PRR is the percentage of required independent DRC-clean pad-to-pad connections. Missing outputs remain in both routability denominators. Higher is better for PR, NRR, and PRR; lower is better for violations, length, and runtime. Rankings use each model's highest NRR score.

Filter:
Model Size Experiment Routability Connectivity Design Rules Routing Quality Efficiency
NRR
(%) ↑		PRR
(%) ↑		 Open
(%) ↓		PShort
(%) ↓		NShort
(%) ↓		 Total
↓		Clr.
↓		Dist.
↓		Polygon
(%)		 TWL
↓		#Vias#Layers PR
↑		RT
↓

Table 2 - Agentic Evaluation of LMMs on PCB Routing

Each model is evaluated across No Tools, Visualization, Routing Score, Semantic Input, Prerouting/Remake, and All Tools. PR uses 500 expected inputs. Avg. Steps is the mean number of recorded agent iterations. Higher is better for PR, NRR, and PRR; lower is better for violations, length, and runtime. Rankings use each model's highest NRR score.

Filter:
Model Size Agent Configuration Routability Connectivity Design Rules Routing Quality Efficiency
NRR
(%) ↑		PRR
(%) ↑		 Open
(%) ↓		PShort
(%) ↓		NShort
(%) ↓		 Total
↓		Clr.
↓		Dist.
↓		Polygon
(%)		 TWL
↓		#Vias#Layers PR
↑		RT
↓		Avg.
Steps

Open, physical-short, and logical-short percentages count each affected net once. Polygon is not a violation. A dash indicates an unavailable result in the manuscript.

About the benchmark

From geometric paths to electrically valid boards

Overview of the OmniRouting benchmark with representative PCB routing cases
Figure 2. Overview of the OmniRouting benchmark with representative cases.

Recent large language models have made progress in constraint-aware navigation and spatial reasoning, but reliable PCB routing additionally demands strict geometric, topological, electrical, and manufacturing constraints. OmniRouting introduces the first large-scale benchmark for this setting, pairing real-world board designs with engineer-verified reference routings and structured evaluation protocols. Current models continue to show weak path planning, poor design-rule adherence, and inconsistent preservation of electrical functionality.

1,681PCB designs
2–8routing layers
109.9Kcomponents
245.4Kpads and SMDs
219.8Kannotated nets
01

Topological path planning

Construct complete routing topologies that connect circuit nets.

02

Design-rule reasoning

Respect clearance, trace-width, via, obstacle, and boundary constraints.

03

Electrical functionality

Preserve schematic intent and reference routing behavior.

04

Agentic tool use

Iteratively refine routes with visualization, scoring, and semantic feedback.

Citation

Reference

T. Lu, K. Lin, Z. Dong, et al., “OmniRouting: A Semantic-Coupled Multimodal Benchmark for Constraint-Aware Spatial Reasoning in PCB Routing,” arXiv preprint arXiv:2608.04434, 2026.

View on arXiv
BibTeX 
@misc{lu2026omnirouting,
  title={OmniRouting: A Semantic-Coupled Multimodal Benchmark for Constraint-Aware Spatial Reasoning in PCB Routing},
  author={Lu, Taiting and Lin, Kaiyuan and Dong, Ziwei and Bei, Sisong and Ye, Haolin and Tian, Yuxin and Liu, Runze and Wang, Mingjia and Zeng, Jingying and Pan, Hongxing and Zhang, Kai and Wang, Haoyu and Shi, Guoliang and Ma, Ling and Yang, Yifan and Lu, Jiaying and He, Qi and Chen, Yi-Chao and Chen, Sung-Liang and Jin, Yincheng and Gowda, Mahanth},
  year={2026},
  eprint={2608.04434},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.04434}
}