RS-EoT-7B: Remote Sensing Evidence-of-Thought

🌐 Project Website | 💻 GitHub Repository | 📄 Paper (ArXiv) | 🤗 Dataset (RS-EoT-4K)

📖 Introduction

RS-EoT-7B is a multimodal reasoning model tailored for Remote Sensing (RS) imagery. It introduces the Evidence-of-Thought (EoT) paradigm to mitigate the "Glance Effect"—a phenomenon where models hallucinate reasoning without genuinely inspecting visual evidence.

Unlike standard VLMs that rely on a single coarse perception, RS-EoT-7B employs an iterative evidence-seeking mechanism. It has been trained using a two-stage pipeline:

1. SFT Cold-Start: Supervised Fine-Tuning on the RS-EoT-4K dataset (synthesized via SocraticAgent) to instill the iterative reasoning pattern.
2. Progressive RL: Reinforcement Learning on Fine-grained Grounding and General VQA tasks to enhance evidence-seeking capabilities and generalize to broader scenarios.

🛠️ Quick Start

Installation

Ensure you have the latest transformers and qwen-vl-utils installed:

pip install transformers
pip install qwen-vl-utils

1. Visual Question Answering (VQA)

This example demonstrates how to ask the model a question and receive a reasoning-backed answer.

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

# Load model and processor
model_name = "ShaoRun/RS-EoT-7B"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_name, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_name)

# Define input image (assumes demo.jpg is in the current directory)
image_path = "./demo.jpg" 

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image_path},
            {"type": "text", "text": "How many cars in this image?"},
        ],
    }
]

# 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
generated_ids = model.generate(**inputs, max_new_tokens=4096)
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[0])

2. Visual Grounding with Visualization

This example shows how to perform visual grounding and visualize the output bounding boxes.

import re
import torch
from PIL import Image, ImageDraw, ImageFont
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

# --- Helper Functions for Parsing and Visualization ---

def extract_bbox_list_in(text: str) -> list[list[float]]:
    """Extracts bounding boxes from the model output text."""
    boxes = []
    text = re.sub(r'\\([{}\[\]":,])', r'\1', text)
    # Pattern to find lists of numbers like [x1, y1, x2, y2]
    pattern = re.compile(r'\[\s*(.*?)\s*\]', flags=re.IGNORECASE | re.DOTALL)
    matches = pattern.findall(text)
    
    number_pattern = r'-?\d+\.\d+|-?\d+'
    for match in matches:
        nums = re.findall(number_pattern, match)
        if len(nums) >= 4:
            # Take the first 4 numbers as the box
            box = [float(num) for num in nums[:4]]
            boxes.append(box)
    return boxes

def visualize_bboxes(img: Image.Image, boxes: list[list[float]], color=(0, 255, 0), width=3) -> Image.Image:
    """Draws bounding boxes on the image."""
    out = img.copy()
    draw = ImageDraw.Draw(out)
    W, H = img.size
    
    for b in boxes:
        if len(b) < 4: continue
        x1, y1, x2, y2 = b[:4]
        
        # Ensure coordinates are within bounds
        x1, y1 = max(0, min(W-1, x1)), max(0, min(H-1, y1))
        x2, y2 = max(0, min(W-1, x2)), max(0, min(H-1, y2))
        
        # Draw rectangle with thickness
        draw.rectangle([x1, y1, x2, y2], outline=color, width=width)
        
    return out

# --- Main Inference Code ---

model_name = "ShaoRun/RS-EoT-7B"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_name, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_name)

# Load Image
image_path = "./demo.jpg"
image = Image.open(image_path).convert('RGB')

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": 'Locate the black car parked on the right in the remote sensing image. Return the coordinates as "[x1, y1, x2, y2]".'},
        ],
    }
]

# Process Inputs
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")

# Generate
generated_ids = model.generate(**inputs, max_new_tokens=4096)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
response = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0]

print(f"Model Response:\n{response}")

# Parse and Visualize
answer_part = response.split("</think>")[-1] 
detection = extract_bbox_list_in(answer_part)

if detection:
    print(f"Detected BBoxes: {detection}")
    vis_img = visualize_bboxes(image, detection)
    vis_img.save("./res.jpg")
    print("Visualization saved to ./res.jpg")
else:
    print("No bounding boxes detected in the response.")

🖊️ Citation

If you use this model in your research, please cite our paper:

@article{shao2025asking,
  title={Asking like Socrates: Socrates helps VLMs understand remote sensing images},
  author={Shao, Run and Li, Ziyu and Zhang, Zhaoyang and Xu, Linrui and He, Xinran and Yuan, Hongyuan and He, Bolei and Dai, Yongxing and Yan, Yiming and Chen, Yijun and others},
  journal={arXiv preprint arXiv:2511.22396},
  year={2025}
}