Research map / Vision-language and multimodal / Multimodal LLMs
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
Cheolhong Min, Jaeyun Jung, Daeun Lee, Hyeonseong Jeon, Yu Su, Jonathan Tremblay, Chan Hee Song, Jaesik Park
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 |
|---|---|---|---|
| evaluate counter-heuristic vs consistent depth examples (no change to model) | Molmo, NVILA, Qwen2.5-VL, Qwen3-VL · EmbSpatial-Bench | accuracy: Qwen2.5-VL +2M: 60.9 -> 24.0 (consistent -> counter) | Hurt |
| scale spatial fine-tuning data 80k -> 2M (ablation) | NVILA-Lite-2B · EmbSpatial-Bench counter | accuracy: 15.5 -> 41.1 | Improved |
| scale spatial fine-tuning data 80k -> 2M (ablation) | Qwen2.5-VL-3B · EmbSpatial-Bench counter | accuracy: 30.2 -> 24.0 | Hurt |
| scale spatial fine-tuning data 80k -> 2M (ablation) | Molmo-7B · EmbSpatial-Bench counter | accuracy: 29.5 -> 39.5 (400k: 27.1) | Mixed |
| scale spatial fine-tuning data to 2M (ablation) | Qwen2.5-VL-3B · BLINK Spatial Relation | accuracy: 83.9 -> 78.3 | Hurt |
| spatial fine-tuning data scaling (ablation) | Molmo, NVILA, Qwen · five spatial benchmarks | accuracy: NVILA 2M: 93.8 CV-3D Depth vs 62.9 BLINK Spat. Rel. | Mixed |
| very large-scale training (RoboRefer 20M+ QA, Qwen3-VL-235B) | VLMs · SpatialTunnel | accuracy gap: +0.046 and +0.068 gaps; Qwen3 mean 0.908 | Improved |
| enable reasoning (ablation) | GPT-5.2 · SpatialTunnel (exact match) | accuracy: 0.613 -> 0.953 | Improved |
| higher distance-axis coherence | VLM representations · EmbSpatial-Bench | counter accuracy: rho = 0.759 (EmbSpatial), 0.804 (CV-Bench-3D) | Improved |
| fine-tuning on spatial data increases vertical axis coherence (ablation) | Molmo / Qwen · EmbSpatial-Bench probing | vertical coherence: Molmo 0.23 -> 0.57; Qwen 0.29 -> 0.59 | Improved |
| RoboRefer vs NVILA 2M (richer training, lower VD-EI) | NVILA-Lite-2B family · EmbSpatial-Bench counter | accuracy: 41.1 -> 59.7 | Improved |
Closest papers by meaning
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.92)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.92)
- Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas (ICML 2025 · similarity 0.92)
- InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)
- Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.92)
- Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.92)
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.
