ReaLM - AdvDiffVLM: leaderboard

Metric: Attack success rate (%; share of adversarial cases where the model gives an incorrect, task-relevant answer) under AdvDiffVLM (adversarial images generated in diffusion latent space) 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 122B A10B43
2Qwen 3.5 27B44.8
3Gemini 3 Flash45.5
4Qwen 3 VL 32B46.5
5Seed 2.0 Mini46.5
6Kimi K2.546.7
7Qwen 3.6 Flash46.7
8Qwen 3.5 9B46.8
9GPT-4.1 Mini49.6

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