Research map / Vision-language and multimodal / Multimodal LLMs
MultihopSpatial: Multi-hop Compositional Spatial Reasoning Benchmark for Vision-Language Model
Youngwan Lee, Soojin Jang, Yoorhim Cho, Seunghwan Lee, Yong-Ju Lee, Sung Ju Hwang
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 |
|---|---|---|---|
| post-train with GRPO on MultihopSpatial-Train | Qwen3-VL-4B-Instruct · MultihopSpatial | MCQ accuracy (%): 37.8 -> 62.9 | Improved |
| post-train with GRPO on MultihopSpatial-Train | Qwen3-VL-4B-Instruct · MultihopSpatial | Acc@50IoU (%): 31.0 -> 53.8 | Improved |
| post-train with GRPO on MultihopSpatial-Train | Qwen3-VL-4B-Instruct · BLINK | accuracy (%): 82.5 -> 85.3 | Improved |
| post-train with GRPO on MultihopSpatial-Train | Qwen3-VL-4B-Instruct · 3DSRBench | accuracy (%): 56.1 -> 56.3 | No clear effect |
| use MultihopSpatial-Train as VLA backbone training | Qwen3-VL-4B-Instruct · CALVIN ABC-D | avg. completed tasks: 3.75 -> 3.98 | Improved |
| use MultihopSpatial-Train as VLA backbone training | Qwen3-VL-4B-Instruct · LIBERO-10 | success rate (%): 35.8 -> 40.0 | Improved |
| add GRPO with MCQ reward on top of SFT (ablation) | Qwen3-VL-4B-Instruct · MultihopSpatial (5-epoch) | MCQ accuracy (%): 58.1 -> 61.8 | Improved |
| add bounding box reward (GIoU) to GRPO (ablation) | Qwen3-VL-4B-Instruct · MultihopSpatial (5-epoch) | Acc@50IoU (%): 51.6 -> 52.3 | Improved |
| add bounding box reward (GIoU) to GRPO (ablation) | Qwen3-VL-4B-Instruct · MultihopSpatial (5-epoch) | MCQ accuracy (%): 61.8 -> 61.1 | Hurt |
| use GRPO with bbox reward instead of SFT (ablation) | Qwen3-VL-4B-Instruct · MultihopSpatial (5-epoch) | Acc@50IoU (%): 49.0 -> 52.3 | Improved |
| increase reward weight alpha from 1 to 2 (beta=1) (ablation) | Qwen3-VL-4B-Instruct · MultihopSpatial (5-epoch) | MCQ accuracy (%): 61.1 -> 52.5 | Hurt |
| increase reward weight beta from 1 to 2 (alpha=1) (ablation) | Qwen3-VL-4B-Instruct · MultihopSpatial (5-epoch) | Acc@50IoU (%): 52.3 -> 51.1 | Hurt |
Closest papers by meaning
- InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.93)
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.92)
- Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes (ICLR 2026 · similarity 0.92)
- STAR-R1: Multi-View Spatial TrAnsformation Reasoning by Reinforcing Multimodal LLMs (CVPR 2026 · similarity 0.91)
- Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
- Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.91)
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.
