Research map / Vision-language and multimodal / Multimodal LLMs
Spatial Amsan: A Benchmark for Perception-Grounded Spatial Reasoning and Action Evaluation in Egocentric Manipulation
Changsoo Jung, Jack Fitzgerald, Ethan Seefried, Mariah Bradford, Nathaniel Blanchard
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 |
|---|---|---|---|
| go from Task 1 (7 colored blocks) to Task 2 (16 irregular blocks) | GPT-5.2 / Gemini 3 Flash · Spatial Amsan | Q3 contact accuracy (%): 74.9 -> 15.9 (GPT-5.2); 69.5 -> 12.2 (Gemini) | Hurt |
| increase number of ground-truth contacts from 1 to 2+ | API VLMs · Spatial Amsan Task 2 | Q3 exact-match (%): 60.2 -> 0.0 (GPT-5.2) | Hurt |
| increase turn position from 1 to 8 | API VLMs · Spatial Amsan Task 1 | Q4 accuracy (%): 81.3 -> 3.9 (GPT-5.2); 81.3 -> 15.8 (Gemini); 79.3 -> 10.5 (Claude) | Hurt |
| increase turn position from 1 to 8 | API VLMs · Spatial Amsan Task 1 | Q1 accuracy (%): 90.7 -> 90.8 (GPT-5.2) | No clear effect |
| insert a single hand-occlusion frame (ablation) | API VLMs · Spatial Amsan Task 1 | Q2 temporal tracking (%): 95.1 -> 85.1 (GPT-5.2); 88.0 -> 80.2 (Claude) | Hurt |
| insert a single hand-occlusion frame (ablation) | API VLMs · Spatial Amsan Task 1 | Q1/Q3/Q4 accuracy: Q4 49.8 -> 51.2 (GPT-5.2) | No clear effect |
| use newer InternVL3.5-8B instead of InternVL3-8B | local VLM · Spatial Amsan | overall accuracy (%): 29.3 -> 17.6 (Task 1); 17.5 -> 5.8 (Task 2) | Hurt |
| use Qwen3-VL-8B instead of Qwen2.5-VL-7B | local VLM · Spatial Amsan Task 1 | overall accuracy (%): 16.8 -> 28.1 | Improved |
| use reasoning-tuned VideoChat-R1-7B over its Qwen2.5-VL base | local VLM · Spatial Amsan | Q4 accuracy (%): 0.7 -> 0.7 (Task 1) | No clear effect |
| frontier API models vs local open-weight models | VLMs · Spatial Amsan Task 1 | overall accuracy (%): 29.3 -> 77.0 | Improved |
Closest papers by meaning
- SAW-Bench: Learning Situated Awareness in the Real World (ICML 2026 · similarity 0.87)
- EgoThinker: Unveiling Egocentric Reasoning with Spatio-Temporal CoT (NeurIPS 2025 · similarity 0.86)
- Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations (ICLR 2026 · similarity 0.86)
- ConsisVLA-4D: Advancing Spatiotemporal Consistency in Efficient 3D-Perception and 4D-Reasoning for Robotic Manipulation (CVPR 2026 · similarity 0.86)
- From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs (CVPR 2026 · similarity 0.85)
- Seeing Together: Multi-Robot Cooperative Egocentric Spatial Reasoning with Multimodal Large Language Models (NeurIPS 2026 · similarity 0.85)
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.
