hassan-wajid/Spatial-Blind-Spots-in-Vision-Language-Models

license: mit

model_evaluated:

• name: Qwen3-VL-2B-Instruct url: https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct

evaluation_notebook:

• https://www.kaggle.com/code/wajidhassanmoosa/blind-spot-qwen3-2b

evaluation_setup: | The model evaluated in this study is Qwen3-VL-2B-Instruct. Evaluation was conducted using the Hugging Face Transformers library with automatic device mapping (device_map="auto") and "bfloat16" dtype selection.

For each example:

1. The image was provided as part of a structured chat message.
2. A reasoning-oriented spatial question was appended.
3. The processor applied the official chat template.
4. The model generated an open-ended answer.
5. The output was compared against the expected spatial ground truth.

Generation was performed with max_new_tokens=32 and no additional prompt engineering beyond concise response constraints.

Full evaluation code is available at: https://www.kaggle.com/code/wajidhassanmoosa/blind-spot-qwen3-2b

blind_spot_description: | While Qwen3-VL-2B-Instruct performs strongly on descriptive multimodal tasks, it exhibits consistent spatial reasoning blind spots under controlled evaluation.

Observed systematic errors include:

1. Dynamic State Tracking Failures: The model fails to correctly update object relationships after hypothetical manipulations (e.g., swapping, lifting, sliding).

2. Reflection Geometry Confusion: The model misinterprets mirrored or perspective-based spatial relationships, particularly in side-view mirrors and reflective surfaces.

3. Mental Rotation and Horizontal Flip Errors: When asked to imagine flipped or rotated configurations, the model often preserves original spatial relationships incorrectly.

4. Containment and Physical Support Failures: The model confuses inside/outside relationships and support structures, suggesting limited internal physical simulation.

5. Small Object Counting Inaccuracy: The model frequently miscounts small identical objects under minimal clutter.

These patterns indicate that large-scale web pretraining prioritizes caption alignment over structured spatial reasoning.

fine_tuning_recommendation: | To mitigate these blind spots, fine-tuning should focus on structured spatial reasoning supervision rather than generic caption datasets.

The required dataset should emphasize:

1. Counterfactual Spatial Manipulation:

   • Object swaps
   • Hypothetical repositioning
   • Horizontal/vertical flips
   • Step-wise state transitions

2. Physically Grounded Containment:

   • Inside/outside reasoning
   • Support/contact relationships
   • Object dependency chains

3. Reflection-Aware Reasoning:

   • Mirror perspective inversion
   • Transparent surfaces
   • Occluded reflections

4. Controlled Counting Tasks:

   • Identical object sets
   • Progressive object addition/removal
   • Varying background complexity

Crucially, prompts must require reasoning, not simple recognition or captioning.

dataset_construction_strategy: | A scalable training corpus can be assembled through a hybrid strategy:

(A) Synthetic Scene Generation (Primary Recommendation) - Use Blender or programmatic 2D engines. - Control object positions precisely. - Automatically compute ground-truth spatial relationships. - Generate transformation variants (flip, swap, move).

(B) Procedural Augmentation - Apply geometric transformations to base scenes. - Automatically recompute spatial labels. - Ensure category-balanced sampling.

(C) Curated Real-World Collection - Mirror-based scenes - Transparent containers - Support/stacking configurations - Annotated with structured reasoning prompts

Synthetic generation is preferred because: - It ensures exact spatial supervision. - It reduces annotation noise. - It enables scalable controlled counterfactual variations.

estimated_dataset_size: | To meaningfully improve spatial reasoning performance:

• 5,000 examples: Minimal measurable improvement in isolated categories.

• 20,000–50,000 examples: Moderate robustness and improved generalization across transformations.

• 100,000+ examples: Strong generalization across diverse spatial compositions, especially if category-balanced.

Each failure category should contain at least 5,000–10,000 diverse examples to avoid shortcut learning.

---