EgoMind: Activating Spatial Cognition through Linguistic Reasoning in MLLMs

Zhenghao Chen^1,2, Huiqun Wang^1,2, Di Huang^1,2✉
1State Key Laboratory of Complex and Critical Software Environment, Beihang University
²School of Computer Science and Engineering, Beihang University

✨ News

• [2026.04.07] 🎉🎉 We have released the model weights and the evaluation code!
• [2026.04.01] 🎉 We have released our paper on arXiv!
• [2026.02.21] 🎉 Our paper has been accepted to CVPR 2026!

🚀 Framework

EgoMind is a Chain-of-Thought (CoT) framework that enables geometry-free spatial reasoning through two key components:

• Role-Play Caption (RPC): Simulates an agent navigating an environment from a first-person perspective, generating coherent descriptions of frame-wise observations and viewpoint transitions to build a consistent global understanding of the scene.
• Progressive Spatial Analysis (PSA): First localizes objects explicitly mentioned in the query, then expands its attention to surrounding entities, and finally reasons about their spatial relationships in an integrated manner.

With only 5K auto-generated SFT samples and 20K RL samples, EgoMind achieves competitive results on VSI-Bench, SPAR-Bench, SITE-Bench, and SPBench, demonstrating the potential of linguistic reasoning for spatial cognition.

🏆 Main Results

EgoMind achieves competitive performance among open-source MLLMs across four spatial reasoning benchmarks, using only 25K training samples (5K CoT-supervised + 20K RL) without any explicit 3D priors.

🔬 Evaluation

1. Environment Installation

# Create and activate a Conda environment (Python 3.11)
conda create -n egomind python=3.11 -y
conda activate egomind

# Install uv, PyTorch, and project dependencies
pip install uv
uv pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
uv pip install -r requirements.txt

2. Model Preparation

Download the model weights into the repo’s models/ directory (from the EgoMind repository root). Requires Hugging Face CLI (pip install huggingface_hub).

huggingface-cli download Hyggge/EgoMind-7B --resume-download --local-dir ./models/EgoMind-7B

After this, point --model_path to models/EgoMind-7B for local inference, or keep using Hyggge/EgoMind-7B to load from the Hub.

3. Dataset Preparation

Download the benchmark data and place them under evaluation/datasets/. See evaluation/datasets/README.md for detailed instructions.

The expected directory structure:

evaluation/datasets/
├── VSI-Bench/
│   ├── qa_processed.jsonl
│   └── data/                  # arkitscenes/, scannet/, scannetpp/
├── SPAR-Bench/
│   ├── qa_processed.jsonl
│   └── data/                  # images/
├── SITE-Bench/
│   ├── qa_processed.jsonl
│   └── data/                  # ActivityNet/, MLVU/, MVBench/, ...
└── SPBench/
    ├── qa_processed.jsonl
    └── data/                  # SPBench-MV-images/, SPBench-SI-images/

4. Running Evaluation

All benchmarks share the same entry point evaluation/run_eval.py. Below are the commands for each benchmark.

VSI-Bench

python evaluation/run_eval.py \
    --model_path models/EgoMind-7B \
    --output_path outputs/EgoMind-7B_vsibench.jsonl \
    --benchmark vsibench

SPAR-Bench

python evaluation/run_eval.py \
    --model_path models/EgoMind-7B \
    --output_path outputs/EgoMind-7B_sparbench.jsonl \
    --benchmark sparbench

SITE-Bench

python evaluation/run_eval.py \
    --model_path models/EgoMind-7B \
    --output_path outputs/EgoMind-7B_sitebench.jsonl \
    --benchmark sitebench

SPBench

python evaluation/run_eval.py \
    --model_path models/EgoMind-7B \
    --output_path outputs/EgoMind-7B_spbench.jsonl \
    --benchmark spbench

Calculate the metric using existing outputs only (skip inference):

python evaluation/run_eval.py \
    --output_path outputs/EgoMind-7B_vsibench.jsonl \
    --benchmark vsibench \
    --only_eval

📜 Citation

If you find our work helpful, please consider citing our paper:

@misc{chen2026egomind,
      title={EgoMind: Activating Spatial Cognition through Linguistic Reasoning in MLLMs}, 
      author={Zhenghao Chen and Huiqun Wang and Di Huang},
      year={2026},
      eprint={2604.03318},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.03318}, 
}