ReaLM - FigStep: leaderboard
ReaLM (2026) is a red-teaming benchmark for physical-world vision-language models on 832 driving, manipulation, grasping, physics, egocentric, diagnostic and scene cases, attacked black-box toward scenario-specific failure targets. This board is the attack success rate under FigStep, misleading text rendered as an auxiliary image: the percent of attacked cases answered incorrectly; lower is better.
Metric: Attack success rate (%; share of adversarial cases where the model gives an incorrect, task-relevant answer) under FigStep (misleading text rendered as an auxiliary image) on the 832 ReaLM physical-world cases (driving, manipulation, grasping, physics, egocentric, diagnostics and scene QA; image, question and ground-truth answer), black-box one-shot evaluation at temperature 0, a Qwen3-8B judge extracting the answer; lower is better. Source: arxiv.org. Saturation forecast: Not forecast. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 43.6 | #164 |
| 2 | GPT-4.1 Mini | 45.4 | #779 |
| 3 | Qwen 3.5 9B | 47 | #803 |
| 4 | Seed 2.0 Mini | 47.2 | #508 |
| 5 | Kimi K2.5 | 48.3 | #256 |
| 6 | Qwen 3.6 Flash | 49.2 | #424 |
| 7 | Qwen 3.5 122B A10B | 49.3 | #331 |
| 8 | Qwen 3.5 27B | 49.5 | #359 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=realm-figstep · How It Works · Data refreshed daily, snapshot 2026-11-13.