EgoPoint-Bench - Real-World Adversarial Resilience: leaderboard

EgoPoint-Bench (2026) asks multimodal models about the object a first-person hand points at, on held-out simulation scenes and real smart-glasses recordings, zero-shot; true/false and single-choice answers by exact match and open answers scored by a GPT-4o judge. This board covers the Adversarial Resilience questions (counterfactual, fallacious or void references) among the 1,162 real-world test questions. Accuracy in percent; higher is better.

Metric: Accuracy (%) on the Adversarial Resilience questions among the 1,162 real-world smart-glasses test questions; zero-shot (direct inference); true/false and single-choice questions by exact match, open-ended answers scored 0 to 1 by a GPT-4o judge; greedy decoding for open models, temperature 0 for API models; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.2 Instant71.27#309
2LLaVA-v1.5-7B69.37#7642
3Gemini 3 Flash63.02#180
4Gemini 3 Pro60.16#146
5Qwen 3 VL 32B Instruct56.83#607
6LLaVA-NeXT-7B56.19#5045
7Qwen 3 VL 8B Instruct53.17#1043
8GPT-5 Mini52.38#373
9InternVL3.5-8B-Instruct50.79#1911
10InternVL3.5-14B-Instruct45.24#1747
11InternVL3.5-2B-Instruct43.65#6576
12GLM-4.6V-Flash43.49#1049

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=egopoint-bench-real-world-adversarial-resilience · How It Works · Data refreshed daily, snapshot 2026-11-22.