CoV: Chain-of-View Prompting for Spatial Reasoning

This is the code repository for the paper:

CoV: Chain-of-View Prompting for Spatial Reasoning

Haoyu Zhao*, Akide Liu*, Zeyu Zhang*, Weijie Wang*, Feng Chen, Ruihan Zhu, Gholamreza Haffari and Bohan Zhuang†

*Equal contribution. †Corresponding author.

ACL 2026 Findings

[arXiv] [HF Paper] [Website] [Trending on X]

Overview

Embodied question answering (EQA) in 3D environments often requires collecting context that is distributed across multiple viewpoints and partially occluded. However, most recent vision--language models (VLMs) are constrained to a fixed and finite set of input views, which limits their ability to acquire question-relevant context at inference time and hinders complex spatial reasoning. We propose Chain-of-View (CoV) prompting, a training-free, test-time reasoning framework that transforms a VLM into an active viewpoint reasoner through a coarse-to-fine exploration process. CoV first employs a View Selection agent to filter redundant frames and identify question-aligned anchor views. It then performs fine-grained view adjustment by interleaving iterative reasoning with discrete camera actions, obtaining new observations from the underlying 3D scene representation until sufficient context is gathered or a step budget is reached.

Updates

• 2026-01-09 We release paper on arXiv.

Project Structure

.
├── cov/                      # Main package
├── scripts/                  # Utility scripts
├── tools/                    # Data processing tools
├── main.py                   # Main entry point
├── pixi.toml                 # Pixi environment configuration
└── README.md

Installation

Prerequisites

• Python 3.9+
• CUDA support (recommended for HabTat Sim)

Using Pixi

The project uses Pixi for dependency management:

# Install dependencies
pixi install

# Activate the environment
pixi shell

How to Run

Setup Environment Variables

Create a .env file in the root directory with your API credentials:

# OpenAI
OPENAI_API_KEY=[your_key_here]

# OpenRouter
OPENROUTER_API_BASE=https://openrouter.api.com/api/v1
OPENROUTER_API_KEY=[your_key_here]

# Dashscop
DASHSCOPE_API_BASE=https://dashscope.aliyuncs.com/compatible-mode/v1
DASHSCOPE_API_KEY=[your_key_here]

Prepare Dataset

1. Download the OpenEQA dataset following the original dataset.
2. Place question files in data/ directory
3. Place scene frames in data/frames/ directory

Run experiment

Run the agent on OpenEQA questions:

# Specify models.
python main.py model=qwen

# Specify min_action_step
python main.py model=qwen min_action_step=7

Custom Models

You can set your own model backend in cov/config.py.

Output

Results are saved to the configured output directory with:

• JSON files containing answers and metadata
• HTML reports showing navigation history and visualizations
• Screenshots of selected views and bird's eye views

Run evaluation

For evaluation, please follow the LLM-Match protocol from OpenEQA.

Citation

If you use Chain of View in your research, please cite this work.

@article{zhao2026cov,
  title={CoV: Chain-of-View Prompting for Spatial Reasoning},
  author={Zhao, Haoyu and Liu, Akide and Zhang, Zeyu and Wang, Weijie and Chen, Feng and Zhu, Ruihan and Haffari, Gholamreza and Zhuang, Bohan},
  journal={arXiv preprint arXiv:2601.05172},
  year={2026}
}