Visual Question Answering on SEED-Bench (Accuracy)
94.5AccuracyDDP
Evaluation Results
| Method | Links | |
|---|---|---|
| DDPbackbone=Gemini-3-Pro2026.04 | 94.5 | |
| Gemini-3-pro2026.04 | 87.2 | |
| GPT-4o2026.04 | 77.1 | |
| Gemini-1.5-pro2026.04 | 77.1 | |
| GRIPModel=Qwen2.5-VL-7B2026.06 | 71.4 | |
| ViTModel=Qwen2.5-VL-7B2026.06 | 70.2 | |
| DINOModel=Qwen2.5-VL-7B2026.06 | 69.8 | |
| Zero ShotModel=Qwen2.5-VL-7B2026.06 | 69.5 | |
| CLIPModel=Qwen2.5-VL-7B2026.06 | 69.2 | |
| GRIPModel=Idefics2-8B2026.06 | 68.2 | |
| ViTModel=Idefics2-8B2026.06 | 67.4 | |
| CLIPModel=Idefics2-8B2026.06 | 66.6 | |
| Zero ShotModel=Idefics2-8B2026.06 | 66.3 | |
| LLaVA-v1.52026.04 | 66.1 | |
| DINOModel=Idefics2-8B2026.06 | 65.4 | |
| Qwen-VL2026.04 | 62.3 | |
| HyCoCLIP-Btrain samples (M)=20.5, trainable params (M)=149.6, Evaluation protocol=Zero-shot, Training=Trained from scratch2026.04 | 46.3 | |
| HAC-S w/ bias tuningtrain samples (M)=13.7, trainable params (M)=0.6, Backbone=CLIP-S2026.04 | 45.6 | |
| HAC-S w/ LN tuningtrain samples (M)=13.7, trainable params (M)=0.5, Backbone=CLIP-S2026.04 | 45.5 | |
| HAC-S w/ seq. adaptertrain samples (M)=13.7, trainable params (M)=1.1, Backbone=CLIP-S2026.04 | 45.1 | |
| HAC-B w/ par. adaptertrain samples (M)=13.7, trainable params (M)=1.2, Evaluation protocol=Zero-shot, Tuning method=par. adapter2026.04 | 45.1 | |
| HAC-S w/ par. adaptertrain samples (M)=13.7, trainable params (M)=1.1, Backbone=CLIP-S2026.04 | 44.9 | |
| HAC-B w/ seq. adaptertrain samples (M)=13.7, trainable params (M)=1.2, Evaluation protocol=Zero-shot, Tuning method=seq. adapter2026.04 | 44.9 | |
| HAC-B w/ LoRAtrain samples (M)=13.7, trainable params (M)=8.0, Evaluation protocol=Zero-shot, Tuning method=LoRA2026.04 | 44.9 | |
| CLIP-Strain samples (M)=20.5, trainable params (M)=85.3, Backbone=CLIP-S2026.04 | 44.7 | |
| CLIP-Btrain samples (M)=20.5, trainable params (M)=149.6, Evaluation protocol=Zero-shot, Training=Trained from scratch2026.04 | 44.7 | |
| HAC-S w/ LoRAtrain samples (M)=13.7, trainable params (M)=11.5, Backbone=CLIP-S2026.04 | 44.6 | |
| HAC-B w/ LN tuningtrain samples (M)=13.7, trainable params (M)=0.7, Evaluation protocol=Zero-shot, Tuning method=LN tuning2026.04 | 44.5 | |
| HAC-B w/ bias tuningtrain samples (M)=13.7, trainable params (M)=0.8, Evaluation protocol=Zero-shot, Tuning method=bias tuning2026.04 | 44.4 | |
| HyCoCLIP-Strain samples (M)=20.5, trainable params (M)=85.3, Backbone=CLIP-S2026.04 | 43.9 | |
| MERU-Btrain samples (M)=12.0, trainable params (M)=149.6, Evaluation protocol=Zero-shot, Training=Trained from scratch2026.04 | 38 | |
| MERU-Strain samples (M)=12.0, trainable params (M)=85.3, Backbone=CLIP-S2026.04 | 37.6 |