SYNCR - Object Counting: leaderboard

Accuracy (%) on counting distinct objects seen across several videos in SYNCR (2026), a synthetic benchmark of multi-video questions with programmatically verified answers; zero-shot four-option multiple choice, exact match, 200 questions per task (chance 25). Higher is better.

Metric: Exact-match accuracy (%) on the SYNCR Object Counting task (counting distinct objects seen across several videos), zero-shot four-option multiple choice over 2 to 4 labelled input videos, exact match, deterministic decoding, on the paper's balanced subset of 200 questions per task; chance is 25; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.468#135
2InternVL3.5-38B66.5#1021
3LLaVA-OV-72B65#1879
4Qwen 3 VL 8B Instruct60.5#1029
5InternVL3.5-14B58.5#1504
6Qwen 3 VL 4B Instruct56#1332
7Qwen 3 VL 32B Instruct54.5#606
8InternVL3.5-8B51.5#1604
9InternVL3.5-4B51.5#1930
10InternVL3.5-2B50.5#3284
11Qwen 2.5 VL 32B Instruct50#1062
12Qwen 2.5 VL 7B Instruct48.5#1902
13Gemini 3 Flash48#175
14Qwen 2.5 VL 72B Instruct46.5#841
15Qwen 3 VL 2B Instruct46.5#2367

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

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