Share your thoughts, 1 month free Claude Pro on usSee more

Visual Question Answering on countbenchqa

93.2Accuracy

RLR³

Updated 5d ago

Evaluation Results

Method	Links
RLR³ 2026.05		93.2
RLR³ 2026.05		92.6
Cont-Squeeze (128 -> 1) 2026.02		92.59
In-Squeeze (128 ... -> 1) 2026.02		92.59
Cont-Squeeze (128 -> 1) 2026.02		92.49
In-Squeeze (128 ... -> 1) 2026.02		92.22
Direct Fine-tuning (r=1) 2026.02		91.94
Direct Fine-tuning (r=1) 2026.02		91.84
Direct Fine-tuning (r=1) 2026.02		91.47
Cont-Squeeze (128 -> 1) 2026.02		91.2
GPT-5 mini 2026.05		91
RLR³ 2026.05		90.3
Official thinking 2026.05		90
Official instruct 2026.05		89.8
RLVR 2026.05		89.7
Base instruct 2026.05		88.7
Gemini-2.0-Flash 2025.12		88.6
RLVR 2026.05		88.3
Gemma 3 4B IT 2026.02		87.4
RLVR 2026.05		87.3
GPT-4o 2025.12		84.7
GPT-5 mini 2026.05		84.1
Qwen2.5-VL-7B 2025.12		82.5
Claude-3.5-Haiku 2025.12		77.6
MRoPE 2026.03		75.97
VALOR 2025.12		75.9
VALOR 2025.12		75.9
Vanilla RoPE 2026.03		75.15
MRoPE-I 2026.03		74.95
ViGoRL 2025.12		74.5
Llama3.2-11B 2025.12		71.5
Vanilla RoPE 2026.03		70.26
GRIT 2025.12		68.6
MRoPE-I 2026.03		67.62
VALOR-RL 2025.12		67.6
VALOR-RL 2025.12		67.6
MRoPE 2026.03		66.6