GeoThinker-F removes VGGT from inference and speeds up spatial reasoning
GeoThinker-F learns to generate geometric and short-term semantic features internally, removing the need to run the external VGGT geometry encoder during inference. The research team reports that latency fell from 1.45 seconds to 0.40 seconds while spatial-reasoning performance remained broadly comparable. In simulation, it achieved a 97.1% task success rate in LIBERO and an 83.3 PDMS score in NAVSIM.
The reported design reduces the inference cost of vision-language spatial reasoning while preserving much of its benchmark performance, which is relevant to latency-sensitive robotics and autonomous-driving research.
导语:不再实时调用外部三维几何编码器,视觉语言模型空间推理提速约 3.6 倍。

不再实时调用外部三维几何编码器,视觉语言模型空间推理提速约 3.6 倍。
作者丨张岂萍
编辑丨幸丽娟
给视觉语言模型多塞一份三维几何信息,它就一定看得更准?
#GeoThinker#GeoThinker-F#VGGT#Sun Yat-sen University#XPeng#Nanyang Technological University#MBZUAI#Qwen2.5-VL-7B#Qwen3-VL-8B#LIBERO#NAVSIM#ReCogDrive