ReaLM - M-Attack: leaderboard

Metric: Attack success rate (%; share of adversarial cases where the model gives an incorrect, task-relevant answer) under M-Attack (cosine-similarity feature matching against a CLIP surrogate ensemble, L-inf 16/255) 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: Around 2029. 13 models tracked.

Top models

#ModelScore
1Gemini 3 Flash43.2
2Kimi K2.543.9
3Seed 2.0 Mini44.6
4Qwen 3.5 27B44.8
5GPT-4.1 Mini44.9
6Qwen 3 VL 32B46.1
7Qwen 3.5 9B46.2
8Qwen 3.5 122B A10B46.3
9Qwen 3.6 Flash48.1

Interactive version: theaggregate.ai/benchmark?slug=realm-m-attack · How It Works · Data refreshed daily, snapshot 2026-09-29.