LAD-Bench: leaderboard
LAD-Bench (2026) tests whether vision-language models notice logical anomalies in everyday scenes, such as a mailbox in the middle of a road or a dolphin with gills, across 1,007 synthetic images; a grading model checks each free-text answer against the annotated anomaly. The board is decay-weighted accuracy in percent: full credit for finding the anomaly unprompted, two thirds after being told something is wrong, one third after a category hint; higher is better.
Metric: Decay-weighted accuracy (%): an anomaly found with no hint scores 100, after being told the image has an anomaly 66.7, after a category hint 33.3, otherwise 0; LAD-Bench: 1,007 synthetic images (gpt-image-1) with one logical anomaly each (nature, residential, urban and collaborative scenes); free-text answers graded by GPT-5-nano against the annotated anomaly; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 70.11 | #161 |
| 2 | Gemini 3 Flash (Preview) | 67.33 | #144 |
| 3 | GPT-5 | 67.2 | #167 |
| 4 | GPT-5 Mini | 65.96 | #348 |
| 5 | Grok 4.1 Fast (Reasoning) | 55.28 | #430 (Grok 4.1 Fast) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lad-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.