See&Trek: Training-Free Spatial Prompting for Multimodal Large Language Models

Official codebase for See&Trek: Training-Free Spatial Prompting for Multimodal Large Language Models

🎉 Our paper has been accepted to NeurIPS 2025.

See&Trek is a training-free and GPU-free spatial prompting framework designed to fundamentally enhance the spatial reasoning capabilities of Multimodal Large Language Models (MLLMs). It addresses two key bottlenecks in existing MLLMs: (1) visual homogeneity caused by uniform frame sampling, and (2) unknown motion due to missing ego-motion cues.

See&Trek achieves this by (i) extracting semantic-rich keyframes using off-the-shelf perception models and (ii) reconstructing camera trajectories via Visual Odometry to annotate keyframes with explicit motion information. Without any model modification or fine-tuning, See&Trek injects structured spatial–temporal priors into MLLMs through a single forward pass, leading to robust improvements across spatial reasoning tasks.

• 🚫 Training-free & GPU-free spatial prompting
• 🔌 Plug-and-play for all open-source and commercial MLLMs
• ⚡ Single-forward inference with zero architecture changes
• 🎥 Semantic-rich keyframe selection + reconstructed motion cues
• 📈 Consistent gains on VSI-Bench and STI-Bench

Follow the official setup for VSI-Benchmark: https://github.com/vision-x-nyu/thinking-in-space

conda create --name vsibench python=3.10
conda activate vsibench

git clone git@github.com:vision-x-nyu/thinking-in-space.git
cd thinking-in-space

git submodule update --init --recursive

cd transformers && pip install -e . && cd ..

pip install -e .
pip install s2wrapper@git+https://github.com/bfshi/scaling_on_scales
pip install deepspeed

Core dependencies:

transformers==4.52.0
lmms_eval==0.2.3
torch==2.6.0
torchvision==0.21.0

Install Ultralytics YOLO:

pip install ultralytics

Make sure dataset paths (e.g., directory, save_dir) in tools/multi-proc-video-skip.py are correctly configured.

Run:

python ./tools/multi-proc-video-skip.py

Outputs are saved to:

./tools/SeeTrek/modified_dataset_yolov8n_2hz_every_4_frames

We also provide preprocessed prior semantics on Google Drive: 🔗 https://drive.google.com/drive/folders/1X3ZpaO8H59H-OxoEturHi2AwWHpm8FMu?usp=drive_link

Place model checkpoints (e.g., InternVL3-Series) into checkpoint-local and update paths in evaluate.sh.

Then run:

bash evaluate.sh --model all --num_processes 2

--num_processes controls how many GPUs to use.

If you find See&Trek useful, please cite:

@article{li2025see,
  title={See\&Trek: Training-Free Spatial Prompting for Multimodal Large Language Model},
  author={Li, Pengteng and Song, Pinhao and Li, Wuyang and Guo, Weiyu and Yao, Huizai and Xu, Yijie and Liu, Dugang and Xiong, Hui},
  journal={arXiv preprint arXiv:2509.16087},
  year={2025}
}