ReaLM - Chain-of-Attack: leaderboard

Metric: Attack success rate (%; share of adversarial cases where the model gives an incorrect, task-relevant answer) under Chain-of-Attack (perturbations co-optimized with re-generated captions) 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 2031. 13 models tracked.

Top models

#ModelScore
1Qwen 3.5 27B51.6
2GPT-4.1 Mini52.9
3Qwen 3.5 9B53.7
4Seed 2.0 Mini53.7
5Qwen 3.6 Flash54.3
6Qwen 3.5 122B A10B54.5
7Gemini 3 Flash54.7
8Qwen 3 VL 32B55.3
9Kimi K2.556.3

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