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

High-Resolution Visual Reasoning on HR-Bench 8K

93.5Accuracy

S1-VL-32B-RL

Updated 23d ago

Evaluation Results

Method	Links
S1-VL-32B-RL 2026.04		93.5
S1-VL-32B-SFT 2026.04		85.1
Gemini 2.5 Pro 2026.04		81.5
Qwen3-VL-235B-A22B-Thinking 2026.04		80.4
Skywork-R1V4-30B 2026.04		79.8
GazeVLM (Ours) 2026.05		78
Qwen3-VL-32B-Thinking 2026.04		77
GazeVLM (w/o gaze bias) 2026.05		74.4
Qwen3.5-4B-Inst 2026.05		74
GPT-5 2026.04		73.75
Gemini 2.5 Flash 2026.04		73.7
Qwen3-VL-4B 2026.05		73.6
DeepEyes 2026.05		72.6
Thyme-VL (7B) 2026.04		72
MGPO 2025.07		71.2
Ground-R1-7B 2026.05		71.1
Qwen2.5-VL-32B 2026.04		70.4
InternVL3-8B 2026.04		69.3
SPARC-4B 2026.05		68.4
MGPO 2025.07		68.1
Pixel-Reasoner-7B 2026.05		66.9
Intern-S1 (235B+6B) 2026.04		66.38
GRPO 2025.07		66.3
Qwen2.5-VL-7B 2026.04		65.3
GRPO 2025.07		64.9
Qwen2.5-VL-7B 2026.05		59.5
GPT-4o 2026.05		55.5
Intern-S1-mini (8B) 2026.04		54.38