(CVPR 2026)Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning
Yuhong Liu
*
Beichen Zhang
*
Yuhang Zang+
*
Yuhang Cao
*
Long Xing
Xiaoyi Dong
*
Haodong Duan
*
Dahua Lin
*
Jiaqi Wang+
+Corresponding authors.
Paper | Homepage | Spatial-SSRL-7B Model | Spatial-SSRL-3B Model | Spatial-SSRL-Qwen3VL-4B Model | Spatial-SSRL-81k Dataset | Daily Paper
News
- [2026/02/25] We have released the Spatial-SSRL-3B Model, initialized from Qwen2.5-VL-3B-Instruct.
- [2026/02/21] Our work has been accepted by CVPR 2026.
- [2025/11/24] We have released the Spatial-SSRL-Qwen3VL-4B Model, initialized from Qwen3-VL-4B-Instruct.
- [2025/11/03] Now you can try out Spatial-SSRL-7B on Spatial-SSRL Space.
- [2025/11/03] We have released the Spatial-SSRL-7B Model, and Spatial-SSRL-81k Dataset.
- [2025/11/02] We have released the Spatial-SSRL repository.
Overview
We are thrilled to introduce Spatial-SSRL, a novel self-supervised RL paradigm aimed at enhancing LVLM spatial understanding. By optimizing Qwen2.5-VL-7B with Spatial-SSRL, the model exhibits stronger spatial intelligence across seven spatial understanding benchmarks in both image and video settings.
Spatial-SSRL is a lightweight tool-free framework that is natually compatible with the RLVR training paradigm and easy to extend to a multitude of pretext tasks. Five tasks are currently formulated in the framework, requiring only ordinary RGB and RGB-D images. And we welcome you to join Spatial-SSRL with effective pretext tasks to further strengthen the capabilities of LVLMs!
Highlights
- Highly Scalable: Spatial-SSRL uses ordinary raw RGB and RGB-D images instead of richly-annotated public datasets or manual labels for data curation, making it highly scalable.
- Cost-effective: Avoiding the need for human labels or API calls for general LVLMs throughout the entire pipeline endows Spatial-SSRL with cost-effectiveness.
- Lightweight: Prior approaches for spatial understanding heavily rely on annotation of external tools, incurring inherent errors in training data and additional cost. In constrast, Spatial-SSRL is completely tool-free and can easily be extended to more self-supervised tasks.
- Naturally Verifiable: Intrinsic supervisory signals determined by pretext objectives are naturally verifiable, aligning Spatial-SSRL well with the RLVR paradigm.
Results
We train Qwen2.5-VL-3B and Qwen2.5-VL-7B with our Spatial-SSRL paradigm and the experimental results across seven spatial understanding benchmarks are shown below.
Quick Start
To directly experience Spatial-SSRL-7B, you can try it out on Spatial-SSRL Space!
Here we provide a code snippet for you to start a simple trial of Spatial-SSRL-7B on your own device. You can download the model from Spatial-SSRL-7B Model before your trial!
from qwen_vl_utils import process_vision_info
model_path = "internlm/Spatial-SSRL-7B" #You can change it to your own local path if deployed already
img_path = "examples/eg1.jpg"
question = "Consider the real-world 3D locations of the objects. Which object has a higher location? A. yellow bear kite B. building"
#We recommend using the format prompt to make the inference consistent with training
format_prompt = "\n You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
model_path, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_path)
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": img_path,
},
{"type": "text", "text": question + format_prompt},
],
}
]
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")
generated_ids = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
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("Model Response:", output_text)
Here we provide a code snippet for you to start a simple trial of Spatial-SSRL-Qwen3VL-4B on your own device. You can download the model from Spatial-SSRL-Qwen3VL-4B Model before your trial!
from qwen_vl_utils import process_vision_info #0.0.14
import torch
model_path = "internlm/Spatial-SSRL-Qwen3VL-4B" #You can change it to your own local path if deployed already
#Change the path of the input image
img_path = "examples/eg_qwen3vl.jpg"
#Change your question here
question = "Question: Consider the real-world 3D locations and orientations of the objects. If I stand at the man's position facing where it is facing, is the menu on the left or right of me?\nOptions:\nA. on the left\nB. on the right\n"
question += "Please select the correct answer from the options above. \n"
#We recommend using the format prompt to make the inference consistent with training
format_prompt = "You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within
model = AutoModelForImageTextToText.from_pretrained(
model_path, torch_dtype=torch.float16, device_map='auto', attn_implementation='flash_attention_2'
)
processor = AutoProcessor.from_pretrained(model_path)
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": img_path,
},
{"type": "text", "text": question + format_prompt},
],
}
]
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")
generated_ids = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
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("Model Response:", output_text[0])
Evaluation
Prepare your environment:
conda create -n spatialssrl python==3.10
conda activate spatialssrl
cd Spatial-SSRL/evaluation
pip install -r requirements.txt
# Recommended
pip install flash-attn --no-build-isolation
Start your evaluation by referring to the tutorials in Eval.md
Todo
- Release the training code.
Cases
Citation
If you find this project useful, please kindly cite:
@article{liu2025spatial,
title={Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning},
author={Liu, Yuhong and Zhang, Beichen and Zang, Yuhang and Cao, Yuhang and Xing, Long and Dong, Xiaoyi and Duan, Haodong and Lin, Dahua and Wang, Jiaqi},
journal={arXiv preprint arXiv:2510.27606},
year={2025}
}
License
Usage and License Notices: The data and code are intended and licensed for research use only.
Acknowledgement
We extend our sincere gratitude to VLMEvalkit, the powerful toolkit to evaluate a vast range of LMMs!