Researchers from The Chinese University of Hong Kong, led by Liang Xiaodan, have developed GeoThinker and GeoThinker-F, two advancements in spatial reasoning for visual language models. GeoThinker improves how models select and utilize external geometric information, reducing interference and boosting performance on benchmarks like VSI-Bench. GeoThinker-F further optimizes this by enabling the model to internally generate spatial representations, eliminating the need for real-time external geometric encoders and resulting in a 3.6x speed increase in spatial reasoning with largely preserved accuracy. AI
IMPACT Accelerates AI's ability to understand and interact with 3D environments, crucial for robotics and autonomous systems.
RANK_REASON The cluster details research findings and model developments (GeoThinker and GeoThinker-F) presented in academic contexts (ICML) and published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
- GeoThinker
- GeoThinker-F
- International Conference on Machine Learning
- Liang Xiaodan
- Mohamed bin Zayed University of Artificial Intelligence
- Qwen2.5-VL-7B
- Qwen3-VL-8B
- The Chinese University of Hong Kong
- VGGT
- VSI-Bench
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
