BenchMarker - Shortcut Detection: leaderboard
Metric: Cohen's kappa (-1 to 1) between the judge and human labels on 271 multiple-choice questions for answer-choice shortcuts: the judge decides whether the question that GPT-5, Gemini 2.5 Pro and Claude 4.5 Sonnet infer from the choices alone fails to match the real question; validation items were sampled from the training splits of 12 NLP multiple-choice datasets, stratified by GPT-5's own BenchMarker predictions (up to ten items predicted flawed and ten predicted not flawed per dataset), then labeled by a researcher with a second annotator checking agreement; default decoding, JSON verdict; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 0.61 | #91 |
| 2 | Claude Sonnet 4.5 | 0.59 | #138 |
| 3 | Claude Haiku 4.5 (20251001) | 0.55 | #251 |
| 4 | GPT-5 Mini (2025-08-07) | 0.53 | #165 |
| 5 | Gemini 2.5 Flash Lite | 0.51 | #413 |
| 6 | Gemma 3 27B (IT) | 0.5 | #509 |
| 7 | c4ai-command-r-08-2024 | 0.5 | #907 |
| 8 | Qwen 3 32B | 0.48 | #424 |
| 9 | Command-R+ (08-2024) | 0.46 | #806 |
| 10 | Gemini 2.5 Pro | 0.43 | #145 |
| 11 | Gemini 2.5 Flash | 0.41 | #237 |
| 12 | Qwen 3 0.6B | 0.39 | #1440 |
| 13 | GPT-5 Nano | 0.38 | #415 |
| 14 | Gemma 3 12B (IT) | 0.31 | #655 |
| 15 | Qwen 3 1.7B | 0.31 | #1186 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=benchmarker-shortcut-detection · How It Works · Data refreshed daily, snapshot 2026-10-11.