Visual Question Answering on VQA lite v2
80AccuracyVanilla
Evaluation Results
| Method | Links | |
|---|---|---|
| VanillaAverage Token Reduction=0.0%, Backbone=Qwen3-VL-8B2026.06 | 80 | |
| VanillaSource=Qwen’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=0%2026.06 | 79.6 | |
| VanillaSource=Qwen’25, Token Reduction=0%, Backbone=Qwen2.5-VL-7B2026.06 | 78.6 | |
| PDrop+RerouteSource=Ours, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.06 | 77.6 | |
| SECONDModel=LLaVA-Next (CLIP-336), LLM=Vicuna-7B, CD=✓2025.06 | 77.5 | |
| FastV+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.06 | 77.2 | |
| SECONDModel=LLaVA-Next (CLIP-336), LLM=Vicuna-7B, CD=✗2025.06 | 76.5 | |
| baselineModel=LLaVA-Next (CLIP-336), LLM=Vicuna-7B, CD=✗2025.06 | 76.4 | |
| FastV+RerouteSource=Ours, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.06 | 76.3 | |
| PDropSource=CVPR’25, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.06 | 75.8 | |
| PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.06 | 75.4 | |
| SECONDModel=LLaVA-OneVision (SigLIP-384), LLM=Qwen2-0.5B, CD=✓2025.06 | 75.1 | |
| PDrop+RerouteAverage Token Reduction=66.7%, Backbone=Qwen3-VL-8B2026.06 | 75.1 | |
| baselineModel=LLaVA-OneVision (SigLIP-384), LLM=Qwen2-0.5B, CD=✗2025.06 | 74.6 | |
| SECONDModel=LLaVA-Next (CLIP-336), LLM=Mistral-7B, CD=✓2025.06 | 74.5 | |
| PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.06 | 74.4 | |
| PDropAverage Token Reduction=66.7%, Backbone=Qwen3-VL-8B2026.06 | 74.2 | |
| PDrop+RerouteSource=Ours, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.06 | 74.1 | |
| SECONDModel=LLaVA-Next (CLIP-336), LLM=Mistral-7B, CD=✗2025.06 | 73.6 | |
| SECONDModel=LLaVA-OneVision (SigLIP-384), LLM=Qwen2-0.5B, CD=✗2025.06 | 73.6 | |
| VCDModel=LLaVA-Next (CLIP-336), LLM=Vicuna-7B, CD=✓2025.06 | 72.9 | |
| FastVSource=ECCV’24, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.06 | 72.9 | |
| FastV+RerouteSource=Ours, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.06 | 72.4 | |
| FastV+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.06 | 72.3 | |
| baselineModel=LLaVA-Next (CLIP-336), LLM=Mistral-7B, CD=✗2025.06 | 72 | |
| FastV+RerouteAverage Token Reduction=66.7%, Backbone=Qwen3-VL-8B2026.06 | 71.8 | |
| PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=88.9%2026.06 | 70.8 | |
| FastVAverage Token Reduction=66.7%, Backbone=Qwen3-VL-8B2026.06 | 70.6 | |
| VCDModel=LLaVA-Next (CLIP-336), LLM=Mistral-7B, CD=✓2025.06 | 70.1 | |
| PDrop+RerouteAverage Token Reduction=77.8%, Backbone=Qwen3-VL-8B2026.06 | 69.3 | |
| PDrop+RerouteSource=Ours, Token Reduction=88.9%, Backbone=Qwen2.5-VL-7B2026.06 | 68.6 | |
| PDropSource=CVPR’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.06 | 68.6 | |
| PDropSource=CVPR’25, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.06 | 67.6 | |
| SECONDModel=Yi-VL (CLIP-448), LLM=Yi-6B, CD=✓2025.06 | 65.3 | |
| FastV+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=88.9%2026.06 | 64.5 | |
| baselineModel=Yi-VL (CLIP-448), LLM=Yi-6B, CD=✗2025.06 | 64.3 | |
| FastVSource=ECCV’24, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.06 | 64.2 | |
| SECONDModel=Yi-VL (CLIP-448), LLM=Yi-6B, CD=✗2025.06 | 63.6 | |
| FastVAverage Token Reduction=77.8%, Backbone=Qwen3-VL-8B2026.06 | 63.4 | |
| FastVSource=ECCV’24, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.06 | 62.8 | |
| PDropAverage Token Reduction=77.8%, Backbone=Qwen3-VL-8B2026.06 | 62.5 | |
| FastV+RerouteAverage Token Reduction=77.8%, Backbone=Qwen3-VL-8B2026.06 | 62.1 | |
| VCDModel=Yi-VL (CLIP-448), LLM=Yi-6B, CD=✓2025.06 | 61.9 | |
| PDrop+RerouteAverage Token Reduction=88.9%, Backbone=Qwen3-VL-8B2026.06 | 60.6 | |
| PDropSource=CVPR’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.06 | 58.6 | |
| PDropSource=CVPR’25, Token Reduction=88.9%, Backbone=Qwen2.5-VL-7B2026.06 | 58.2 | |
| VCDModel=LLaVA-OneVision (SigLIP-384), LLM=Qwen2-0.5B, CD=✓2025.06 | 55 | |
| FastVSource=ECCV’24, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.06 | 53.1 | |
| FastV+RerouteSource=Ours, Token Reduction=88.9%, Backbone=Qwen2.5-VL-7B2026.06 | 52.3 | |
| PDropAverage Token Reduction=88.9%, Backbone=Qwen3-VL-8B2026.06 | 51.6 | |
| PDropSource=CVPR’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=88.9%2026.06 | 47.3 | |
| FastV+RerouteAverage Token Reduction=88.9%, Backbone=Qwen3-VL-8B2026.06 | 42.7 | |
| FastVAverage Token Reduction=88.9%, Backbone=Qwen3-VL-8B2026.06 | 40.8 | |
| FastVSource=ECCV’24, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=88.9%2026.06 | 38.9 | |
| FastVSource=ECCV’24, Token Reduction=88.9%, Backbone=Qwen2.5-VL-7B2026.06 | 37.7 |