Research map / Vision-language and multimodal / Multimodal LLMs
OmniCoT: A Benchmark for Global and Multi-Step Panoramic Reasoning
Haocong He, Chenfei Liao, Zichen Wen, Zihao Dongfang, Xu Zheng, Bin Ren, Chang Su, Zixin Zhang, Harold Haodong Chen, Hongfei Zhang, Weijia Li, Kailun Yang and 4 more
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 |
|---|---|---|---|
| add SFT on OmniCoT-T to Qwen2.5-VL-7B (ablation) | Qwen2.5-VL-7B-Instruct · OmniCoT-B | overall accuracy: 23.53 -> 49.31 | Improved |
| add GRPO stage after SFT (ablation) | OmniCoT-R1 · OmniCoT-B | overall accuracy: 49.31 -> 59.54 | Improved |
| add GRPO after SFT (ablation) | OmniCoT-R1 · OmniCoT-B Locate | accuracy: 39.44 -> 51.82 | Improved |
| add GRPO after SFT (ablation) | OmniCoT-R1 · OmniCoT-B | CoT F1: 38.07 -> 51.14 | Improved |
| use CoT instead of direct answer for open-source Qwen2.5-VL-72B (ablation) | Qwen2.5-VL-72B-Instruct · OmniCoT-B | overall accuracy: 26.51 -> 35.26 (+8.75) | Improved |
| use CoT instead of direct answer for LLaVA-v1.5-7B (ablation) | LLaVA-v1.5-7B · OmniCoT-B | overall accuracy: 4.91 -> 13.77 (+8.86) | Improved |
| use CoT instead of direct answer for LLaVA-v1.5-13B (ablation) | LLaVA-v1.5-13B · OmniCoT-B | overall accuracy: 17.38 -> 15.03 (-2.35) | Hurt |
| use CoT instead of direct answer for Gemini3-Flash (ablation) | Gemini3-Flash · OmniCoT-B | overall accuracy: 39.97 -> 28.44 (-11.53) | Hurt |
| use CoT instead of direct answer for GPT-5 (ablation) | GPT-5 · OmniCoT-B | overall accuracy: 42.66 -> 39.72 (-2.94) | Hurt |
| increase spatial anchors from 0 to 10 (ablation) | InternVL3.5-14B · OmniCoT-B | overall accuracy: 28.35 -> 34.23 | Improved |
| increase spatial anchors from 0 to 10 (ablation) | GPT-5 · OmniCoT-B | overall accuracy: 40.60 -> 44.03 | Improved |
| increase anchors from 0 to 3 (default) (ablation) | Qwen2.5-VL-7B-Instruct · OmniCoT-B | overall accuracy: 25.79 -> 23.53 | Hurt |
Closest papers by meaning
- PanoWorld: Towards Spatial Supersensing in 360◦ Panorama World (NeurIPS 2026 · similarity 0.92)
- More than the Sum: Panorama-Language Models for Adverse Omni-Scenes (CVPR 2026 · similarity 0.90)
- MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.90)
- OmniSpatial: Towards Comprehensive Spatial Reasoning Benchmark for Vision Language Models (ICLR 2026 · similarity 0.90)
- From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models (ICML 2026 · similarity 0.89)
- InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · 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.
