---

Overview

Vision-Language Encoders (VLEs) are widely adopted as the backbone of zero-shot referring image segmentation (RIS), enabling text-guided localization without task-specific training. Yet the conventionally used final-layer multimodal embeddings prioritize global semantic alignment — and the biases latent in mid-layer representations have remained largely unexplored.

Through layer-wise investigation, we surface two coupled consequences of this over-reliance on final-layer features:

• Vision: embeddings exhibit weak sensitivity to positional cues, limiting spatial grounding.
• Language: multilingual text embeddings form language-dependent geometric shifts within the shared representation space.

💡 Contributions

• Layer-wise bias analysis — reveals that final-layer VLE representations suppress positional sensitivity and exhibit language-dependent geometric drift in multilingual settings.
• Mid-layer spatial map — identifies a spatial grounding pathway in VLE mid-layers applicable for zero-shot RIS without task-specific training.
• Mixed-language mid-layer grounding — leveraging cross-lingual representation diversity at mid-layers yields strong spatial grounding gains (+7–8 mIoU) and improved zero-shot retrieval.

---

Installation

apt-get install -y libgl1

uv pip install spacy git+https://github.com/openai/CLIP.git h5py gem_torch open_clip_torch==2.24.0 pydantic==2.9.1 opencv-python-headless

uv pip install torch==2.6.0+cu118 torchvision==0.21.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

cd ./third_party/old_detectron2
uv pip install -e . --no-build-isolation

uv pip install pillow==9.5.0

uv pip install spacy==3.7.6 

python -m spacy download en_core_web_lg

Checkpoint: Mask2Former

---

Dataset Structure

Biased2Grounded/
├── assets/
│   ├── multilingual_meta_data/
│   ├── refer_data/
│   ├── ├── images/
│   ├── ├── refcoco/
│   ├── ├── refcoco+/
│   ├── ├── refcocog/

---

Experiments

# Raw Spatial Maps for CLIP, SigLIP, and SigLIP-2
notebooks/*.ipynb

# run spatial grounding evaluation
bash infer.sh
bash infer_baseline.sh 
bash infer_pmaponly.sh # p-map only ablation

# run multilingual evaluation
bash run.sh
bash run_baseline.sh
bash run_comparison.sh # final-layer centroid

---

Citation

If you use this work in your research, please cite:

@misc{an2026blindpositionbiasedlanguage,
      title={Blind to Position, Biased in Language: Probing Mid-Layer Representational Bias in Vision-Language Encoders for Zero-Shot Language-Grounded Spatial Understanding}, 
      author={Na Min An and Inha Kang and Minhyun Lee and Hyunjung Shim},
      year={2026},
      eprint={2509.23098},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.23098}, 
}