Research map / Vision-language and multimodal / Multimodal LLMs
Anchored, Not Graded: How Vision-Language Models Fail at Slant-from-Texture Perception
Qian Zhang, Michal Golovanevsky, Fulvio Domini, James Tompkin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune with SFT (3, 10, 100 epochs) | Qwen2.5-VL-3B · Slant test | MAE (degrees): 48.3 -> 18.9 -> 5.6 -> 4.7 | Improved |
| Fine-tune with SFT (100 epochs) | Qwen2.5-VL-3B · Slant test | peak anchor count: 208 -> 10 | Improved |
| Fine-tune with SFT | Qwen2.5-VL-3B · Curvature sign | accuracy (%): 50 -> 96.8 | Improved |
| Vary prompt type (detail, style, JSON) (ablation) | Zero-shot VLMs · Slant test | slant error: p=0.157 | No clear effect |
| Add in-context labeled examples (ablation) | Zero-shot VLMs · Slant test | slant error: Median slant errors remain above 40 | No clear effect |
| Lower temperature or repeat queries and average (ablation) | Zero-shot VLMs · Slant test | slant error | No clear effect |
| Mean-ablate single attention heads in LM (ablation) | Qwen2.5-VL-3B · Slant test | anchoring | No clear effect |
| Mean-ablate single attention heads in LM (ablation) | Qwen2-VL-7B · Slant test | anchoring | No clear effect |
| Increase FOV and optical slant | Qwen2.5-VL-3B after SFT · Curvature sign | sign flip errors | Improved |
| Use final layer instead of mid layer for probing | Qwen2.5-VL-3B ViT · Probe | R2 physical slant: 0.879 -> 0.769 | Hurt |
| Add patch merger after final layer | Qwen2.5-VL-3B ViT · Probe | R2 FOV: 0.844 -> 0.884 | Improved |
| Final layer of early-fusion model | Chameleon-7B · Probe | R2 physical slant: 0.938 -> 0.065 | Hurt |
Closest papers by meaning
- Can We Talk Models Into Seeing the World Differently? (ICLR 2025 · similarity 0.89)
- See It from My Perspective: How Language Affects Cultural Bias in Image Understanding (ICLR 2025 · similarity 0.88)
- Revisiting Visual Corruptions in LVLMs: A Shape-Texture Perspective on Model Failures (CVPR 2026 · similarity 0.87)
- Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic Saliency (NeurIPS 2026 · similarity 0.87)
- VHELM: A Holistic Evaluation of Vision Language Models (NeurIPS 2024 · similarity 0.87)
- Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
