This model card describes the ViGoRL (Visually Grounded Reinforcement Learning) model, introduced in our paper "Grounded Reinforcement Learning for Visual Reasoning".

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

Model Overview

ViGoRL is a vision-language model fine-tuned using reinforcement learning (RL) to explicitly anchor textual reasoning steps to visual coordinates. Inspired by human visual cognition, ViGoRL employs multi-turn visual grounding, dynamically zooming into image regions to perform fine-grained visual reasoning and grounding.

This model was trained using supervised fine-tuning (SFT) on visually-grounded reasoning traces generated via Monte Carlo Tree Search (MCTS), followed by reinforcement learning with Group Relative Policy Optimization (GRPO).

Model Details

Base Architecture: Qwen2.5-Vision-Language (3B or 7B parameters)
Training Paradigm:
- Supervised Fine-Tuning on MCTS-generated reasoning traces
- Group Relative Policy Optimization (GRPO)
- Multi-turn visual grounding with dynamic zoom-in feedback (if "Multiturn" appears in name)

Use Cases

This model excels in visual reasoning tasks that require precise visual grounding and region-level reasoning. Please see model name for specific domain.

Spatial Reasoning: SAT-2, BLINK, RoboSpatial
Visual Search: V*Bench
Web Interaction and Grounding: ScreenSpot (Pro and V2), VisualWebArena

Usage

You can load this model easily using Hugging Face's Transformers library:

python

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-7b-Web-Action", 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-7b-Web-Action",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

# default processer
processor = AutoProcessor.from_pretrained("gsarch/ViGoRL-7b-Web-Action")

# 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-7b-Web-Action", min_pixels=min_pixels, max_pixels=max_pixels)

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.

Important: This model requires a system prompt for proper usage. Please see the model's chat template for details.

Datasets and Training Data

Training datasets and generated reasoning chains are publicly available:

Citation

If you use ViGoRL in your research or applications, please cite our paper:

bibtex

@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}
}

Contact

For questions, feedback, or collaborations, please reach out to Gabriel Sarch or open an issue in our GitHub repository.