LDU-Bench: leaderboard
LDU-Bench (2026) evaluates multimodal LLMs on lithography defect review over real IC-SEM images, scored deterministically without an LLM judge. The Lithography Closure Score is a power mean (p = 0.5) of the four task scores; higher is better, in percent.
Metric: Lithography Closure Score (%; power mean with p = 0.5 of the four task scores: defect-triage and morphology macro-F1, localization DICU and cause-analysis rubric score; IC-SEM lithography and integrated-circuit review images; one zero-shot run per model at temperature 0 where exposed; invalid outputs stay in the denominator; deterministic scoring, no LLM judge). Source: arxiv.org. Saturation forecast: Around November 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 47.4 | #121 |
| 2 | GLM-5V Turbo | 46.2 | #342 |
| 3 | MiniMax-M3 | 43.9 | #265 |
| 4 | Qwen 3.6 Plus | 40.8 | #217 |
| 5 | Claude Opus 4.6 | 37.8 | #100 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ldu-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.