List your company
ResearchFormerly WizWand

Visual Question Answering on countbenchqa

93.2AccuracyRLR³
65.53672.71879.987.082Dec 9, 2025Jan 6, 2026Feb 3, 2026Mar 4, 2026Apr 1, 2026Apr 29, 2026May 28, 2026
Updated 5mo ago

Evaluation results

MethodLinks
RLR³Training Source=ViRL2026.0593.2
RLR³Training Source=DeepVision2026.0592.6
Cont-Squeeze (128 -> 1)Optimization steps=+200 steps2026.0292.59
In-Squeeze (128 ... -> 1)Optimization mode=Min steps2026.0292.59
Cont-Squeeze (128 -> 1)Optimization steps=+700 steps2026.0292.49
In-Squeeze (128 ... -> 1)Optimization mode=Standard2026.0292.22
Direct Fine-tuning (r=1)Optimization steps=+700 steps2026.0291.94
Direct Fine-tuning (r=1)Optimization steps=+0 steps2026.0291.84
Direct Fine-tuning (r=1)Optimization steps=+200 steps2026.0291.47
Cont-Squeeze (128 -> 1)Optimization steps=+0 steps2026.0291.2
GPT-5 miniVersion=high2026.0591
RLR³Training Source=OpenMMR2026.0590.3
Official thinkingSource=Qwen3-VL2026.0590
Official instructSource=Qwen3-VL2026.0589.8
RLVRTraining Source=OpenMMR2026.0589.7
Base instruct2026.0588.7
Gemini-2.0-FlashModel Category=Proprietary2025.1288.6
RLVRTraining Source=DeepVision2026.0588.3
Gemma 3 4B ITZero-shot=true2026.0287.4
RLVRTraining Source=ViRL2026.0587.3
GPT-4oModel Category=Proprietary2025.1284.7
GPT-5 miniVersion=minimal2026.0584.1
Qwen2.5-VL-7BModel Category=Open-Source2025.1282.5
Claude-3.5-HaikuModel Category=Proprietary2025.1277.6
MRoPEPositional Encoding Method=MRoPE, DIPE Enhancement=+DIPE2026.0375.97
VALORBase Model=Qwen3-8B2025.1275.9
VALORModel Category=Open-Source2025.1275.9
Vanilla RoPEPositional Encoding Method=Vanilla RoPE, DIPE Enhancement=+DIPE2026.0375.15
MRoPE-IPositional Encoding Method=MRoPE-I, DIPE Enhancement=+DIPE2026.0374.95
ViGoRLBase Model=Qwen2.5-VL-7B2025.1274.5
Llama3.2-11BModel Category=Open-Source2025.1271.5
Vanilla RoPEPositional Encoding Method=Vanilla RoPE, DIPE Enhancement=Base2026.0370.26
GRITBase Model=Qwen2.5-VL-3B2025.1268.6
MRoPE-IPositional Encoding Method=MRoPE-I, DIPE Enhancement=Base2026.0367.62
VALOR-RLBase Model=Qwen3-8B2025.1267.6
VALOR-RLModel Category=Open-Source2025.1267.6
MRoPEPositional Encoding Method=MRoPE, DIPE Enhancement=Base2026.0366.6