GeoThinker-F
A spatial intelligence framework that enables a multimodal large language model to predict spatial representations internally and eliminate dependence on external 3D geometry encoders, using a two-stage decoupled training strategy.
Timeline (last 90 days)
GeoThinker-F is a spatial intelligence framework that frees multimodal large language models (MLLMs) from dependence on external 3D encoders, achieving spatial reasoning by having the model intrinsically predict the required spatial representations.
GeoThinker-F's training process is split into two stages: in the first stage the MLLM backbone is frozen and only the lightweight latent frame prediction (LFP) head is trained; in the second stage the external 3D expert is removed and the spatial grounding temporal fusion (SGT) module is jointly optimized.
The SGT module's parameter count is far smaller than that of a complete 3D encoder, which is one of the main sources of the 3.6x speedup.
GeoThinker-F maintains competitive performance on multiple spatial intelligence benchmarks while achieving significant efficiency gains.
GeoThinker-F achieves up to a 3.6x speedup in inference.
GeoThinker-F reduces latency from 1.45 seconds to 0.40 seconds.
GeoThinker-F completely eliminates its dependence on 3D geometry encoders at runtime.
The currently public information comes only from a research abstract and lacks the experimental details of a full paper, such as the specific architectural design of the LFP head, the implementation of the SGT fusion module, the evaluation benchmarks, the baseline comparisons, and the hardware conditions and input scale of the 3.6x speedup.
All Facts (8)
GeoThinker-F is a spatial intelligence framework that frees multimodal large language models (MLLMs) from dependence on external 3D encoders, achieving spatial reasoning by having the model intrinsically predict the required spatial representations.
50%UnverifiedGeoThinker-F's training process is split into two stages: in the first stage the MLLM backbone is frozen and only the lightweight latent frame prediction (LFP) head is trained; in the second stage the external 3D expert is removed and the spatial grounding temporal fusion (SGT) module is jointly optimized.
50%UnverifiedThe SGT module's parameter count is far smaller than that of a complete 3D encoder, which is one of the main sources of the 3.6x speedup.
50%UnverifiedGeoThinker-F maintains competitive performance on multiple spatial intelligence benchmarks while achieving significant efficiency gains.
50%UnverifiedGeoThinker-F achieves up to a 3.6x speedup in inference.
50%UnverifiedGeoThinker-F reduces latency from 1.45 seconds to 0.40 seconds.
50%UnverifiedGeoThinker-F completely eliminates its dependence on 3D geometry encoders at runtime.
50%UnverifiedThe currently public information comes only from a research abstract and lacks the experimental details of a full paper, such as the specific architectural design of the LFP head, the implementation of the SGT fusion module, the evaluation benchmarks, the baseline comparisons, and the hardware conditions and input scale of the 3.6x speedup.
50%