ReaLM - AdvEDM: leaderboard

Metric: Attack success rate (%; share of adversarial cases where the model gives an incorrect, task-relevant answer) under AdvEDM-A (semantic editing that steers attention toward the target region) 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 September 2027. 13 models tracked.

Top models

#ModelScore
1Gemini 3 Flash36.5
2Qwen 3.5 27B38.8
3Qwen 3.5 122B A10B39.3
4Qwen 3.6 Flash40.9
5Kimi K2.541.6
6Qwen 3 VL 32B43.1
7GPT-4.1 Mini43.5
8Qwen 3.5 9B43.9
9Seed 2.0 Mini44.1

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