3DVerifier - Explanation-aware Overall: leaderboard
3DVerifier v1 Table1 full 2,581-example/10-task benchmark; direct inference temperature0/top_p1/max_tokens512. AppendixD.2 API gpt-6-astra-2026-09-03; preserve Gemma NVFP4, Qwen3.5/Qwen3.6/GLM FP8, others local BF16. Static task scores use balanced accuracy. Human400-subset and random-reference rows are not model cells. Motion uses native MP4 for Gemini3.8Flash/Gemma4/Qwen3.5/Qwen3.6 and fixed16-frame contact sheet for GPT6/GPT5.6/InternVL/GLM/OmniVerifier/Qwen2.5; do not pool their input protocols invisibly. ExpOA is explanation-aware task-macro BA: reference-based text-only Qwen3.5-122BA10B FP8 judge, temp0 seed42, max512, thinkingoff; false claims require valid rejection explanation. No subject images supplied to judge. Metric: Explanation-aware balanced accuracy (%). Higher is better. Source: https://arxiv.org/html/2610.22360v1
Metric: Explanation-aware balanced accuracy (%). Source: arxiv.org. Saturation forecast: Around August 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 59.6 | #18 |
| 2 | Gemini 3.8 Flash | 57.8 | #21 |
| 3 | Qwen 2.5 VL 7B Instruct | 22.1 | #649 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=3dverifier-explanation-aware-overall · How It Works · Data refreshed daily, snapshot 2026-10-17.