ReaLM - PA-Attack: leaderboard

Metric: Attack success rate (%; share of adversarial cases where the model gives an incorrect, task-relevant answer) under PA-Attack (untargeted out-of-distribution prototype attack, L-inf 8/255) 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 2028. 13 models tracked.

Top models

#ModelScore
1Qwen 3.5 27B38.1
2Qwen 3.5 122B A10B39.4
3GPT-4.1 Mini40.6
4Qwen 3.6 Flash41.7
5Gemini 3 Flash41.9
6Qwen 3 VL 32B42.3
7Qwen 3.5 9B42.8
8Kimi K2.543
9Seed 2.0 Mini45.4

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