DGP Generalization - ScienceQA Image: leaderboard
Dynamic visual-layer routing systems on original multimodal evaluation tables. Table 2 separately compares Qwen3-VL-2B with SigLIP2-L/Qwen3-1.7B, and LLaVA-NeXT with CLIP-L/Vicuna-7B, against their DGP-trained counterparts. Exact vision and language backbones are pinned. DGP v1 trains CLIP ViT-L/14-336 plus Vicuna-7B or Vicuna-13B on LLaVA-558K pretraining and LLaVA-mixed-665K supervised data, one epoch per stage. The language model and DGP modules train jointly while the vision encoder stays frozen. Native scores cover GQA, ScienceQA Image, VQAv2, VizWiz, SEED All/Image, LLaVA-Wild, MM-Vet, MMBench English/Chinese and POPE. Each DGP row is a separately trained system, never bare Vicuna. Cited historical baseline rows remain source evidence and are not emitted again. Question access: Primary result reviewed; exact scored release availability not confirmed.
Metric: Native benchmark score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DGP Qwen3-VL-2B (Benchmark Sweep Run: source=DGPGeneralization; SigLIP2-L; Qwen3-1.7B; trained DGP) | 86.9 | |
| 2 | Qwen3-VL-2B (Benchmark Sweep Run: source=DGPGeneralization; SigLIP2-L; Qwen3-1.7B; native baseline) | 85.3 | |
| 3 | DGP LLaVA-NeXT Vicuna-7B (Benchmark Sweep Run: source=DGPGeneralization; CLIP-L; trained DGP) | 72.3 | |
| 4 | LLaVA-NeXT Vicuna-7B (Benchmark Sweep Run: source=DGPGeneralization; CLIP-L; native baseline) | 69.5 |
Interactive version: theaggregate.ai/benchmark?slug=dgp-generalization-scienceqa-image · How It Works · Data refreshed daily, snapshot 2026-11-06.