data/              # dataset roots or symlinks
src/               # DeCoDe source code
episode_logs/      # fixed few-shot episode sampling logs and runtime outputs
cache/             # default model cache, e.g. Qwen3-VL-8B-Instruct
configs/           # YAML experiment configs

Create a conda environment and install PyTorch 2.10:

conda create -n decode python=3.11 -y
conda activate decode

pip install "torch==2.10.*"
pip install -r requirements.txt

If your CUDA setup needs a specific PyTorch wheel, install the matching PyTorch 2.10 build first, then install the remaining requirements.

Download or symlink datasets under data/ so they match the expected reader layout. See data/README.md for the dataset-specific folder structure and download notes.

After setting up the environment and data, run:

python src/scripts/run_few_shot.py --config configs/qwen3vl_decompose.yaml --dataset mini

The default config uses Qwen/Qwen3-VL-8B-Instruct, decompose_semantic support/query prompting, and writes prediction JSONL files to episode_logs/runs/<dataset>/. Pairwise support/query comparisons are evaluated in batches controlled by experiment.batch_size or --batch_size.

Prompt modes:

• decompose_semantic: compare support/query images with the support class label in the prompt.
• decompose_anonymous: compare support/query images without class labels. Add --domain_info to either prompt mode to use dataset-specific domain wording, such as aircraft variant, bird species, or action category.

Configured datasets: mini, cub, dogs, arabicsign, yoga, hieroglyph, ucf, aircraft, domain, lego, butterfly, and industrial.

If you find this work useful, please cite:

@misc{wang2026decomposecomparedecidemultimodal,
  title         = {Decompose, Compare, and Decide: Multimodal LLMs are Implicit Few-Shot Learners},
  author        = {Yunhan Wang and Eshika Khandelwal and Edson Araujo and Walid Bousselham and Nina Shvetsova and Hilde Kuehne},
  year          = {2026},
  eprint        = {2607.00125},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.00125}
}