🚀 ViGoRL：用于视觉推理的视觉基础强化学习

ViGoRL（视觉基础强化学习）模型旨在解决视觉推理任务中精确视觉定位和区域级推理的问题。它通过强化学习进行微调，能将文本推理步骤与视觉坐标明确关联，在多种视觉推理任务中表现出色。

🚀 快速开始

你可以使用Hugging Face的Transformers库轻松加载此模型：

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

# # default: Load the model on the available device(s)
# model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
#     "gsarch/ViGoRL-Multiturn-3b-Visual-Search", torch_dtype="auto", device_map="auto"
# ) # replace with any of the ViGoRL models

# We recommend enabling flash_attention_2 for better acceleration and memory saving.
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "gsarch/ViGoRL-Multiturn-3b-Visual-Search",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

# default processer
processor = AutoProcessor.from_pretrained("gsarch/ViGoRL-Multiturn-3b-Visual-Search")

# The default range for the number of visual tokens per image in the model is 4-16384.
# You can set min_pixels and max_pixels according to your needs, such as a token range of 256-1280, to balance performance and cost.
# min_pixels = 256*28*28
# max_pixels = 1280*28*28
# processor = AutoProcessor.from_pretrained("gsarch/ViGoRL-Multiturn-3b-Visual-Search", min_pixels=min_pixels, max_pixels=max_pixels)

# messages = [
#     {
#         "role": "user",
#         "content": [
#             {
#                 "type": "image",
#                 "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
#             },
#             {"type": "text", "text": "What color is the leash."},
#         ],
#     }
# ]

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "path/to/image.png",
            },
            {"type": "text", "text": "QUERY HERE"},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text) # this will output a single tool call turn of the model if version is multiturn.
# Example output of gsarch/ViGoRL-Multiturn-3b-Visual-Search: ['<think> The leash appears to be red, as seen near the dog\'s paw and the person\'s hand. (1028, 1093). </think>\n<tool_call>\n{"name": "search_coordinate", "arguments": {"coordinate": [1028, 1093]}}\n</tool_call>']

⚠️ 重要提示

此模型的正确使用需要系统提示。具体细节请参阅模型的聊天模板。

✨ 主要特性

• 视觉基础强化学习：ViGoRL是一个视觉语言模型，通过强化学习（RL）进行微调，能将文本推理步骤明确锚定到视觉坐标上。
• 多轮视觉定位：受人类视觉认知的启发，ViGoRL采用多轮视觉定位，动态放大图像区域，以执行细粒度的视觉推理和定位。

📚 详细文档

模型概述

ViGoRL是一个视觉语言模型，通过强化学习（RL）进行微调，将文本推理步骤明确锚定到视觉坐标上。受人类视觉认知的启发，ViGoRL采用多轮视觉定位，动态放大图像区域，以执行细粒度的视觉推理和定位。

该模型首先在通过蒙特卡罗树搜索（MCTS）生成的视觉基础推理轨迹上进行监督微调（SFT），然后使用组相对策略优化（GRPO）进行强化学习。

模型详情

使用场景

该模型在需要精确视觉定位和区域级推理的视觉推理任务中表现出色。具体领域请参阅模型名称。

• 空间推理：SAT - 2、BLINK、RoboSpatial
• 视觉搜索：V*Bench
• 网页交互与定位：ScreenSpot（Pro和V2）、VisualWebArena

数据集和训练数据

训练数据集和生成的推理链是公开可用的：

• 代码
• Hugging Face上的ViGoRL数据集

引用

如果你在研究或应用中使用了ViGoRL，请引用我们的论文：

@article{sarch2025vigorl,
    title={Grounded Reinforcement Learning for Visual Reasoning},
    author={Sarch, Gabriel and Saha, Snigdha and Khandelwal, Naitik and Jain, Ayush and Tarr, Michael J and Kumar, Aviral and Fragkiadaki, Katerina},
    year={2025}
}

联系信息

如有问题、反馈或合作需求，请联系Gabriel Sarch，或在我们的GitHub仓库中提出问题。

作者：Gabriel Sarch、Snigdha Saha、Naitik Khandelwal、Ayush Jain、Michael J. Tarr、Aviral Kumar、Katerina Fragkiadaki

本模型卡片描述了ViGoRL（视觉基础强化学习）模型，该模型在我们的论文"Grounded Reinforcement Learning for Visual Reasoning"中被提出。