Research map / Vision-language and multimodal / Multimodal LLMs
ScAle: Attention Head Scaling as a Minimal Adapter for Spatial Reasoning in Vision–Language Models
Rahul Chowdhury, Timothy Rupprecht, Xuan Shen, Pu Zhao, Yanzhi Wang
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 |
|---|---|---|---|
| learn bounded per-head attention scalars on a frozen backbone | LLaVA-7B · SpatialEval Maze-Nav | accuracy: 31.7 -> 68.9 | Improved |
| scale attention heads instead of only MLP outputs (ablation) | ScAle LLaVA-7B · SpatialEval Maze-Nav | accuracy: 44.0 -> 68.9 | Improved |
| add MLP scaling on top of attention scaling (ablation) | ScAle (attn) LLaVA-7B · SpatialEval | accuracy: Maze-Nav 68.9 -> 69.7; Spatial-Map 45.0 -> 45.3 | No clear effect |
| increase training data from 5% to 20% (ablation) | ScAle (attn) LLaVA-7B · SpatialEval Spatial-Grid | accuracy: 51.7 -> 68.6 | Improved |
| combine LoRA on QKV with ScAle scaling | LoRA · SpatialEval Maze-Nav | accuracy: 71.6 -> 78.9 (vs LoRA-all) | Improved |
| apply LoRA only to last layer instead of ScAle attention scaling | LLaVA-7B · SpatialEval Spatial-Grid | accuracy: 68.6 -> 29.2 | Hurt |
| train ScAle on Spatial-Grid and evaluate on real-world VGQA | Qwen2.5-VL-3B · VGQA | accuracy: 84.5 -> 92.2 | Improved |
| train ScAle on spatial task, evaluate on other spatial task | Qwen2.5-VL-3B · Maze-Nav | accuracy: 26.0 -> 27.3 (Spatial-Grid adapter) | No clear effect |
| train ScAle on spatial tasks, evaluate on GLUE/SST2 | Qwen2.5-VL-3B · GLUE/SST2 | accuracy: 65.8 -> 58.8 (Maze-Nav), 88.3 (Spatial-Map) | Mixed |
| apply ScAle to hallucination benchmark | LLaVA-7B · POPE | accuracy: 82.1 -> 88.2 | Improved |
Closest papers by meaning
- Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas (ICML 2025 · similarity 0.92)
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.90)
- SIFThinker: Spatially-Aware Image Focus for Visual Reasoning (AAAI 2026 · similarity 0.90)
- InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.89)
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.
