GeoThinker: Active Spatial Reasoning

• GeoThinker is a spatial reasoning framework that defines active geometry integration by retrieving geometric evidence on-demand rather than through passive fusion.
• It uses Spatial-Grounded Fusion and Importance Gating to selectively integrate geometric features based on semantic cues, reducing redundant signals.
• Empirical results show a peak VSI-Bench score of 72.6 and improved performance in downstream tasks like embodied referring and autonomous driving.

Searching arXiv for the primary GeoThinker paper and closely related spatial-reasoning work. GeoThinker is a framework for spatial reasoning in multimodal LLMs that shifts geometry integration from passive fusion to active perception. Rather than exposing geometric priors as a global stream, it enables the model to selectively retrieve geometric evidence conditioned on its internal reasoning demands, using Spatial-Grounded Fusion at carefully selected VLM layers and Importance Gating to bias per-frame attention toward task-relevant structures. In the reported evaluation, GeoThinker achieves a peak score of 72.6 on VSI-Bench and shows improved spatial perception in downstream scenarios including embodied referring and autonomous driving (Li et al., 5 Feb 2026).

1. Definition and conceptual position

GeoThinker is presented as a framework for active geometry integration in spatial reasoning. Its central claim is that recent multimodal LLMs are strong at semantic recognition but remain weak at reliable spatial reasoning, especially when the task requires understanding object count, absolute and relative distance, relative direction, object size, room size, route planning, approach order, multi-view spatial localization, egocentric movement consequences, or downstream planning and grounding tasks (Li et al., 5 Feb 2026).

The framework is motivated by a critique of prior geometry-aware multimodal systems. Earlier approaches are described as relying on passive fusion, in which geometry is treated either as input or as supervision. In these settings, geometric features are injected into the visual stream through addition, concatenation, or alignment losses, but are still exposed as a uniformly available stream. GeoThinker argues that this induces three recurrent problems: semantic-geometry misalignment, redundant geometry signals, and indiscriminate fusion (Li et al., 5 Feb 2026).

This positioning places GeoThinker within a broader line of work on geometry-aware reasoning, but its distinguishing feature is the treatment of geometry as an on-demand resource rather than a globally fused modality. This suggests a shift from feature mixing toward selective evidence retrieval during reasoning (Li et al., 5 Feb 2026).

2. Core principle: active geometry integration

The defining mechanism of GeoThinker is that the base VLM first produces semantic hidden states from images or video and text, and these hidden states then serve as visual-semantic priors for querying a separate geometry representation. Geometry is therefore not injected everywhere; it is retrieved selectively according to what the model currently needs for the task (Li et al., 5 Feb 2026).

The paper frames this as a move from passive fusion to active perception. In this formulation, spatial reasoning is not improved simply by making more geometric features available. Instead, the model must determine which geometric evidence is relevant, where it is relevant, and at what stage of the reasoning process it should be integrated. The goal is to reduce both irrelevant geometric noise and the mismatch between abstract semantic states and local structural features (Li et al., 5 Feb 2026).

A plausible implication is that GeoThinker treats spatial reasoning as a conditional retrieval problem inside the model: semantic context specifies the demand, and geometry is integrated only where that demand is strongest. This interpretation is consistent with the paper’s emphasis on task-dependent and spatially selective geometry use (Li et al., 5 Feb 2026).

3. Architecture and fusion mechanism

The architectural overview described in the paper is decoupled. A visual encoder or VLM backbone processes RGB frames and text, while a separate geometry encoder, specifically VGGT, extracts geometry-aware features from the same frames. At selected VLM layers, GeoThinker applies Spatial-Grounded Fusion, allowing hidden semantic image states to query geometry through cross-attention, and Importance Gating predicts a localized attention bias from semantic context (Li et al., 5 Feb 2026).

Two architectural components are emphasized.

Spatial-Grounded Fusion: this is applied at carefully selected VLM layers. The paper states that semantic visual priors selectively query and integrate task-relevant geometry via frame-strict cross-attention (Li et al., 5 Feb 2026).

Importance Gating: this calibrates the fusion by biasing per-frame attention toward task-relevant structures. Rather than weighting all geometric inputs equally, the model predicts which frames or structures should matter more for the current reasoning demand (Li et al., 5 Feb 2026).

The stated design principle is that high-level semantic states and geometry should not be fused indiscriminately. GeoThinker instead uses semantic priors to retrieve geometry from the external 3D stream. This makes the geometry pathway conditional on the VLM’s internal reasoning state, which is the paper’s operational definition of active perception (Li et al., 5 Feb 2026).

A plausible implication is that the choice of selected VLM layers is intended to reduce semantic-geometry misalignment by intervening where semantic features remain informative for querying structure but have not yet become too abstract. The paper identifies layer selection as deliberate, though it does not provide a broader formal theory of optimal placement in the supplied material (Li et al., 5 Feb 2026).

4. Relation to broader geometry-aware reasoning research

GeoThinker belongs to a wider movement in multimodal reasoning that seeks stronger geometric grounding. Related work in the supplied literature illustrates several neighboring strategies.

"Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views" proposes 3DThinker, which introduces an internal geometric latent that functions like a mental 3D scene. It does not rely on explicitly labeled 3D data for training and improves multiple spatial benchmarks by aligning reasoning-time latent tokens to a 3D foundation model (Chen et al., 21 Oct 2025). The contrast is instructive: 3DThinker emphasizes internal 3D mentaling, whereas GeoThinker emphasizes selective retrieval from an external geometry stream (Li et al., 5 Feb 2026, Chen et al., 21 Oct 2025).

"Geometric Feature Enhanced Knowledge Graph Embedding and Spatial Reasoning" addresses spatial reasoning in Geospatial Knowledge Graphs by aligning relation embeddings with topology, direction, and distance features. Its contribution is geographically informed representation learning rather than visual-spatial fusion, but it shares the broader aim of injecting explicit spatial structure into learned representations (Hu et al., 2024). This suggests that GeoThinker participates in a larger trend toward making learned systems respect geometry rather than infer it only implicitly.

A broader methodological context is provided by "GeoBench: Rethinking Multimodal Geometric Problem-Solving via Hierarchical Evaluation," which argues that final-answer evaluation obscures failure modes in perception, planning, theorem application, and self-correction (Feng et al., 30 Dec 2025). Although GeoBench addresses geometric problem solving rather than video- or image-based spatial intelligence, its diagnostic stance is relevant: GeoThinker’s contribution can be read as a response to the specific bottleneck of spatial structure integration, not as a complete theory of multimodal reasoning (Li et al., 5 Feb 2026, Feng et al., 30 Dec 2025).

5. Empirical performance and downstream scenarios

The principal quantitative claim reported for GeoThinker is that it achieves a peak score of 72.6 on VSI-Bench (Li et al., 5 Feb 2026). The paper further states that it shows robust generalization and significantly improved spatial perception in downstream scenarios including embodied referring and autonomous driving (Li et al., 5 Feb 2026).

Because the supplied material does not enumerate a full benchmark table, detailed cross-model comparisons beyond the VSI-Bench peak are not available here. The paper nonetheless positions GeoThinker as setting a new state of the art in spatial intelligence and emphasizes robustness across complex downstream scenarios rather than gains on a single narrow task (Li et al., 5 Feb 2026).

This suggests that the framework’s main empirical strength lies not only in benchmark accuracy but in transfer to settings where spatial reasoning must remain grounded under more complex perceptual and decision-making demands. The downstream examples imply that selective geometry retrieval is useful when the model must localize referents or reason over scene structure in embodied or driving contexts (Li et al., 5 Feb 2026).

6. Significance, interpretation, and open implications

GeoThinker’s significance lies in its reformulation of geometry use inside multimodal models. It does not reject geometry encoders; instead, it rejects the assumption that geometry should be fused everywhere and all at once. The paper’s central claim is that the ability to actively integrate spatial structures is essential for next-generation spatial intelligence (Li et al., 5 Feb 2026).

Three broader implications follow from the supplied description.

First, GeoThinker suggests that spatial reasoning benefits from conditional access to structure rather than unconditional access to all structure. This is consistent with the paper’s diagnosis of redundant geometry signals and task-dependent relevance (Li et al., 5 Feb 2026).

Second, it implies that semantic reasoning and geometric reasoning should be coupled asymmetrically: semantics drives the query, geometry provides supporting evidence. This differs from approaches that assume semantic and geometric streams should simply be co-equal inputs at all stages (Li et al., 5 Feb 2026).

Third, GeoThinker strengthens a larger research direction in which multimodal models become more selective, tool-like, and evidence-seeking. A plausible interpretation is that active geometry integration plays for spatial structure a role analogous to retrieval in knowledge-intensive language tasks: rather than storing or fusing everything uniformly, the model learns when and how to access the right information.

The supplied material does not provide a dedicated limitations section, so stronger claims about failure modes, compute cost, or ablation-derived bottlenecks would be speculative. What can be stated directly is that GeoThinker is proposed as a framework-level alternative to passive geometry fusion, anchored by Spatial-Grounded Fusion, Importance Gating, VGGT-based geometry encoding, and empirical gains on VSI-Bench and downstream spatial tasks (Li et al., 5 Feb 2026).

In this sense, GeoThinker marks a conceptual transition in multimodal spatial reasoning: from geometry as a static auxiliary stream to geometry as selectively retrieved evidence, integrated only when the model’s reasoning process calls for it.