E3VS-Bench - Counting: leaderboard

E3VS-Bench (2026) embodied active-perception episodes on Counting questions (instances of an object class across the scene): a vision-language agent moves a 5-DoF camera in 3D Gaussian Splatting scenes, and a GPT-5.1 answerer and judge score the outcome 5 (correct) or 1 (incorrect). Mean score on a 1-5 scale; higher is better.

Metric: Judge score on Counting questions (instances of an object class across the scene) of the E3VS-Bench test split (377 episodes, 21 3D Gaussian Splatting scenes): the model drives a 5-DoF camera agent (at most 25 steps) to a viewpoint from which the question can be answered; GPT-5.1 then answers from the agent's final view and a GPT-5.1 judge scores the answer 5 (correct) or 1 (incorrect); mean judge score on a 1-5 scale; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 2.5 Pro2.04#276
2Gemini 3 Pro1.96#145
3Qwen 3 VL 8B1.96#917
4GPT-5.1Non-reasoning1.88#240
5Gemini 3 FlashNon-reasoning1.88#175
6Qwen 3 VL 30B A3B1.84#784
7Step3 VL 10B1.64#1080
8InternVL3.5-8B1.6#1604

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

Interactive version: theaggregate.ai/benchmark?slug=e3vs-bench-counting · How It Works · Data refreshed daily, snapshot 2026-11-21.