List your company
ResearchFormerly WizWand

Visual Question Answering on HRBench-4K

0.7925AccuracyFOCUS
0.33490.45370.57250.6913Jun 26, 2025Aug 19, 2025Oct 13, 2025Dec 6, 2025Jan 30, 2026Mar 25, 2026May 19, 2026
Updated 5mo ago

Evaluation results

MethodLinks
FOCUSBase Model=Qwen-2.5-VL2025.060.7925
Thyme-7BThinking Level=Thinking with Images2026.020.783
ViLAVT-7BThinking Level=Chatting with Images2026.020.755
DeepEyes-7BThinking Level=Thinking with Images2026.020.751
UG-SearchBase Model=Qwen2.5-VL-7B2025.100.749
Baseline (No Recompute)Model=Qwen3-VL-8B, Recomputation budget (k)=02026.030.7488
Qwen3VL-32B-Instruct2026.050.745
CLAIMDIFF-RL actor-onlyReward=actor-only, Initialization=Qwen2026.050.745
Pixel-Reasoner-7BThinking Level=Thinking with Images2026.020.729
No RecomputeModel=Qwen3-VL-8B, Recomputation budget (k)=22026.030.7275
InfoFlow KVModel=Qwen3-VL-8B, Recomputation budget (k)=22026.030.7263
EPICModel=Qwen3-VL-8B, Recomputation budget (k)=22026.030.725
CacheBlendModel=Qwen3-VL-8B, Recomputation budget (k)=22026.030.7238
Qwen-2.5-VL2025.060.7162
HART-7BResolution=4, 023 × 3, 5032026.020.711
InternVL3-8BThinking Level=Non-Thinking2026.020.708
InternVL3-7BResolution=4, 023 × 3, 5032026.020.708
Qwen2.5-VL-7BResolution=4, 023 × 3, 5032026.020.701
UG-SearchBase Model=LLaVA-OV-7B2025.100.701
Qwen2.5-VL-7BThinking Level=Thinking about Images2026.020.698
InfoFlow KVModel=Qwen3-VL-8B, Recomputation budget (k)=42026.030.6913
No RecomputeModel=Qwen3-VL-8B, Recomputation budget (k)=42026.030.6888
CacheBlendModel=Qwen3-VL-8B, Recomputation budget (k)=42026.030.6863
ZoomEyeBase Model=LLaVA-OV-7B2025.100.684
Qwen2.5-VL-7BThinking Level=Non-Thinking2026.020.678
Qwen 2.5 VLInput Resolution=1792x17922025.110.6775
EPICModel=Qwen3-VL-8B, Recomputation budget (k)=42026.030.6725
ThymeBase Model=Qwen2.5-VL-7B2025.100.666
CLAIMDIFF-RL relativeReward=relative, Initialization=SFT2026.050.665
CropVLMReward=LL, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.110.6638
CropVLMReward=Accuracy, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.110.6575
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=2048, Cropping Strategy=CropVLM2025.110.6513
LLaVA-OV-7BBase Model=LLaVA-OV-7B2025.100.649
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=512, Cropping Strategy=CropVLM2025.110.6475
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=1024, Cropping Strategy=CropVLM2025.110.6475
LLaVA-OneVision-7BThinking Level=Non-Thinking2026.020.643
LLaVA-OneVision-7BResolution=4, 023 × 3, 5032026.020.643
HOLISTIC-RLReference Setting=w/ ref, Initialization=SFT2026.050.64
ViCropBase Model=Qwen2.5-VL-7B2025.100.633
CLAIMDIFF-RL actor-onlyReward=actor-only, Initialization=SFT2026.050.62
TextCoTBase Model=Qwen2.5-VL-7B2025.100.606
VILASR-7BThinking Level=Thinking with Images2026.020.605
Qwen2.5-VL-7BBase Model=Qwen2.5-VL-7B2025.100.601
SpaceR-7BThinking Level=Thinking about Images2026.020.581
HOLISTIC-RLReference Setting=w/o ref, Initialization=SFT2026.050.575
MRoPEPositional Encoding Method=MRoPE, DIPE Enhancement=+DIPE2026.030.545
SFT2026.050.535
MRoPE-IPositional Encoding Method=MRoPE-I, DIPE Enhancement=+DIPE2026.030.5212
Qwen 2.5 VL 3BBase Model=Qwen 2.5 VL 3B, Resolution=-, Cropping Strategy=None2025.110.5188
MRoPE-IPositional Encoding Method=MRoPE-I, DIPE Enhancement=Base2026.030.5162
Vanilla RoPEPositional Encoding Method=Vanilla RoPE, DIPE Enhancement=+DIPE2026.030.51
Vanilla RoPEPositional Encoding Method=Vanilla RoPE, DIPE Enhancement=Base2026.030.5075
MRoPEPositional Encoding Method=MRoPE, DIPE Enhancement=Base2026.030.4763
CropVLMBase Model=LLaVA 1.5 7B, Resolution=1024, Cropping Strategy=CropVLM2025.110.4388
CropVLMBase Model=GPT 4.1 nano, Resolution=2048, Cropping Strategy=CropVLM2025.110.4313
CropVLMBase Model=LLaVA 1.5 7B, Resolution=2048, Cropping Strategy=CropVLM2025.110.4138
CropVLMBase Model=GPT 4.1 nano, Resolution=1024, Cropping Strategy=CropVLM2025.110.4138
CropVLMBase Model=LLaVA 1.5 7B, Resolution=512, Cropping Strategy=CropVLM2025.110.3988
GPT 4.1 nanoBase Model=GPT 4.1 nano, Resolution=-, Cropping Strategy=None2025.110.3875
CropVLMBase Model=GPT 4.1 nano, Resolution=512, Cropping Strategy=CropVLM2025.110.3863
LLaVA 1.5 7BBase Model=LLaVA 1.5 7B, Resolution=-, Cropping Strategy=None2025.110.3525