CrashSight - Robustness: leaderboard

Hallucination probe questions (cannot-determine, absent-entity, temporal-trap and false-premise items) of CrashSight (2026), a multiple-choice VQA benchmark on real roadside-camera traffic crash videos answered zero-shot from four frames. Accuracy in percent; higher is better.

Metric: Accuracy (%) on hallucination-probe questions (cannot-determine, absent-entity, temporal-trap and false-premise items whose answer resists fabrication), four-option multiple-choice VQA on held-out roadside-camera crash videos of CrashSight-VQA (4 frames at 1 fps, zero-shot, deterministic decoding, exact letter match); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1InternVL3-8B82.2#1782
2Qwen 2.5 VL 7B Instruct73.3#1902
3Qwen 2.5 VL 3B Instruct71.1#2713
4InternVL3-2B71.1#3290
5LLaVA-NeXT-Video-7B66.1#5734
6LLaVA-OneVision-0.5B18.9#7956

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=crashsight-robustness · How It Works · Data refreshed daily, snapshot 2026-11-21.