PInVerify (Attribute Decomposition): leaderboard
Active instance verification: deciding whether an object seen from chosen viewpoints matches a fine-grained description, over 3,000 offline episodes of 18 object categories; PInVerify (Jiang, 2026). The multimodal model runs in the paper's multi-view pipeline that checks each attribute of the description separately (Grounding DINO detection, LLM next-best-view, at most six steps). Accuracy from 0 to 1; higher is better.
Metric: Verification accuracy (0-1; share of 3,000 episodes, 1,000 each of matching, same-category and different-category instances, decided correctly; multi-view pipeline with attribute decomposition, LLM next-best-view, Grounding DINO detection, at most six steps). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3-VL-4B | 0.85 | |
| 2 | Qwen3-VL-8B | 0.8 | |
| 3 | SenseNova-SI-1.2-InternVL3-8B | 0.71 |
Interactive version: theaggregate.ai/benchmark?slug=pinverify-attribute-decomposition · How It Works · Data refreshed daily, snapshot 2026-11-12.