ModelsOfficial

中山大学 GeoThinker-F 去掉 3D 编码器,空间推理提速 3.6 倍

Sun Yat-sen University's GeoThinker-F Speeds Up Spatial Reasoning 3.6x by Dropping the 3D Encoder

Why we picked it

中山大学这个思路挺聪明:训练时借 VGGT 学 3D 空间线索,推理时把 3D 编码器整个扔掉,延迟从 1.45 秒压到 0.40 秒。做空间智能体、具身智能部署的可以看看这套蒸馏方法。

中山大学团队发布 GeoThinker-F,空间推理延迟从 1.45 秒降至 0.40 秒,速度提升 3.6 倍。核心思路是在训练阶段让模型从 VGGT(Visual Geometry Grounded Transformer)中学习三维空间线索,推理阶段则完全移除 3D 编码器,不再依赖它。

Body

具体机制上,GeoThinker-F 把 VGGT 的几何知识蒸馏进视觉语言模型,使模型在推理时直接从 2D 图像推断空间关系,省去了逐帧运行 3D 编码器的开销。这解决了传统空间推理管线中 3D 重建模块带来的延迟瓶颈,同时保留了空间理解能力。

对部署空间智能体或具身智能系统的开发者来说,这种"训练时用重模型、推理时轻量化"的蒸馏路线,是降低延迟与算力成本的可复用范式。

Original · pandaily

Sun Yat-sen University's GeoThinker-F Speeds Up Spatial Reasoning 3.6x by Dropping the 3D Encoder

Sun Yat-sen University's GeoThinker-F learns 3D cues from VGGT during training, then drops it, cutting spatial reasoning latency from 1.45 s to 0.40 s.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include GeoThinker-F, VGGT, spatial reasoning, reasoning models.

Primary source verification: published directly by pandaily with first-party authority.