Introduction
Vision Language Models (VLMs) have achieved remarkable success in semantic tasks like image captioning and object recognition. However, they consistently struggle with "spatial intelligence"—the ability to reason about the 3D world. This "flatland" problem stems from two primary architectural bottlenecks: the reliance on 2D patch-level tokens for input and the use of discrete text tokenization for numerical output.
GEODE (Geometric-Output and Decoupled-Input Engine) addresses these limitations by decoupling 3D spatial reasoning into a dedicated input co-processor and a direct numerical regression head. This specialized architecture allows a relatively small 1.5B parameter model to outperform much larger 7B+ models and proprietary systems like GPT-4o on complex spatial benchmarks.
The GEODE Architecture
GEODE introduces two plug-and-play modules designed to bypass the inherent limitations of standard 2D-centric VLMs.
Decoupled Rationale Module (DRM)
Standard VLMs struggle to interpret geometric relationships from 2D pixels alone. The DRM acts as a spatial co-processor, ingesting explicit 3D data such as point clouds and camera poses. Using a cross-attention mechanism, it fuses these 3D features with 2D visual data to generate Rationale Tokens (<spatio>). These tokens distill complex spatial logic into a format the core LLM can readily utilize, providing pre-computed geometric context without overwhelming the main model with raw 3D processing.
Direct Regression Head (DRH)
Predicting precise measurements (e.g., "4.6 meters") is difficult for LLMs because they treat numbers as sequences of discrete text tokens. GEODE implements an "Embedding-as-Value" paradigm. When a numerical answer is required, the model emits a specialized control token (<REG>). The hidden-state embedding of this token is then routed to a lightweight Multi-Layer Perceptron (MLP) that performs continuous regression. This allows the system to output precise coordinates or distances as single mathematical values rather than fragmented strings.
Experimental Results and Key Insights
The effectiveness of GEODE was evaluated using VSI-Bench, a comprehensive benchmark for Video Spatial Intelligence containing over 5,000 QA pairs derived from egocentric videos.
Performance Benchmarks
Despite its modest 1.5B parameter count, GEODE achieved state-of-the-art (SOTA) results across multiple categories:
- Overall Score: GEODE scored 37.0, surpassing GPT-4o (34.0) and the much larger InternVL2-40B (36.0).
- Absolute Distance: In tasks requiring precise measurement, GEODE scored 25.7, nearly doubling the performance of Qwen2.5-VL-7B (14.8).
- Object Counting: The model achieved a score of 61.9, significantly outperforming standard 7B open-source models.
Efficiency through Decoupling
The primary insight of this research is that decoupling is synonymous with efficiency. By offloading geometric processing to the DRM and numerical output to the DRH, the core LLM is liberated to function as a high-level semantic dispatcher. Ablation studies confirmed that removing either module leads to a significant drop in performance, proving that addressing both the input and output bottlenecks is essential for true spatial intelligence.
Limitations and Future Directions
While GEODE represents a significant step forward, it currently relies on external models for initial 3D reconstruction and point cloud encoding, which can limit its overall performance. Additionally, the current regression head uses fixed control tokens for specific tasks.
Future work will focus on expanding the Direct Regression Head to handle more complex, dense outputs, such as 3D semantic segmentation voxels or continuous navigational trajectories for autonomous robotics. This evolution could eventually allow VLMs to move beyond simple question-answering and into active, real-world spatial interaction.
Summary updated