---
license: mit
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen2.5-VL-7B-Instruct
datasets:
- MLL-Lab/viewsuite
---

# Planning with the Views

This repository contains a model checkpoint presented in the paper [Planning with the Views](https://huggingface.co/papers/2605.29563).

[**Project Page**](https://viewagent.github.io) | [**GitHub**](https://github.com/mll-lab-nu/ViewAgent) | [**Paper**](https://arxiv.org/pdf/2605.29563)

## Overview

Can VLMs predict how each camera move changes the view, and plan many such moves ahead? This capability, called **view planning**, requires (1) understanding how a single action transforms the view, and (2) composing many such transformations across multi-turn plans to identify a target view. 

**ViewSuite** is a 3D point-cloud environment and benchmark suite for view planning, built on real ScanNet indoor scenes. It probes view planning through three diagnostic tasks:
- **Path-to-View (P2V)**: Predict the resulting view from an action sequence.
- **View-to-Path (V2P)**: Infer the action sequence between two views.
- **Interactive View Planning (IVP)**: Plan view changes over multiple turns to identify a target view.

This model is an optimized version of **Qwen2.5-VL-7B**, trained using an iterative framework that alternates self-exploration with view graph distillation. This approach significantly closes the planning gap found in frontier VLMs, improving performance on interactive view planning tasks.

## Citation

If you find ViewAgent or these checkpoints useful in your research, please consider citing:

```bibtex
@article{wang2026viewagent,
  title   = {Planning with the Views},
  author  = {Wang, Kangrui and Li, Linjie and Yang, Zhengyuan and Chen, Shiqi and
             Wang, Zihan and Fei-Fei, Li and Wu, Jiajun and Guibas, Leonidas and
             Wang, Lijuan and Li, Manling},
  year    = {2026}
}
```