DRBench (Dense-Scene Reasoning): leaderboard
DRBench (Shi et al., Shanghai Jiao Tong University, 2026, introduced with DRScaffold) asks questions about cluttered indoor (Hypersim) and urban driving (Cityscapes) scenes that require grounding several objects and their relations. Its five categories are multi-hop perception, spatial reasoning, affordance reasoning, anomaly detection and false-premise rejection. Combined accuracy in percent over 959 test questions: multiple-choice answers by exact letter match and open answers by a Qwen-Max semantic judge, weighted by question count; higher is better.
Metric: Combined accuracy (%; question-count-weighted mean of multiple-choice letter accuracy and open-ended semantic accuracy judged by Qwen-Max; 959 test questions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Phi-4 Multimodal Instruct | 45.8 | #3914 |
| 2 | Qwen 2.5 VL 32B Instruct | 44.7 | #1055 |
| 3 | InternVL2.5-2B | 32.5 | #6868 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=drbench-dense-scene-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-12.