Rule2DRC: leaderboard
Design-rule-check script synthesis from Rule2DRC (Kim, Byun, Hwang, Park and Song, ICML 2026): the model turns a natural-language chip design rule into an executable KLayout DRC script, over 1,000 rules (310 from the SkyWater 130nm PDK, 490 synthetic multi-constraint and 200 syntax-coverage rules) with 13,921 labelled evaluation layouts and the 60K-token KLayout API documentation in the prompt. Success rate in percent counts a single sampled script only when its violation output matches the ground-truth script on every layout (mean of 30 runs; higher is better).
Metric: Success rate (%; one sampled DRC script per rule with the 60K-token KLayout DRC API documentation in the prompt, correct when its violation output matches the ground-truth script on every evaluation layout; 1,000 rules, mean of 30 runs). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B (Medium) | 32.5 | #699 (GPT-OSS-120B) |
| 2 | GPT-OSS-20B (Medium) | 16.9 | #1150 (GPT-OSS-20B) |
| 3 | Qwen 3 30B A3B 2507 Instruct | 14.1 | #1078 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rule2drc · How It Works · Data refreshed daily, snapshot 2026-11-12.