Reasoning Over Space (ROS)

• ROS is a framework that formalizes spatial reasoning using symbolic calculi, geometric models, and neuro-symbolic integration to infer spatial relations.
• It enables structured spatial inference in diverse applications such as robotics, geographic decision-making, and vision-language tasks.
• Empirical benchmarks reveal that combining chain-of-thought reasoning with reinforcement learning significantly improves ROS performance over perception-only methods.

Reasoning Over Space (ROS) encompasses the representational frameworks, computational models, and learning paradigms enabling agents—artificial or biological—to perform structured inference about spatial entities, their properties, relationships, and dynamics. ROS is foundational to spatial cognition, robot scene understanding, geographic decision-making, generative recommendation systems, and spatially-aware vision-language inference. It ranges from qualitative calculi and logical reasoning to neuro-symbolic and deep learning methods, addressing the alignment of internal reasoning processes with spatially grounded, verifiable outcomes.

1. Formal Foundations of Spatial Reasoning

ROS systems formalize spatial domains using symbolic calculi, geometric abstractions, or compositional encodings. Foundational frameworks include:

• Qualitative Calculi. Predicate-based representations, such as RCC-8 (mereotopology) or orientation/distance calculi, define a finite set of jointly exhaustive, pairwise disjoint base relations $\mathcal{B}$ over a spatial domain $\mathcal{D}$, with constraint networks assigning relations between each pair (or tuple) of entities (Lee et al., 2022, Schwertfeger, 2019).
• Quantitative and Mixed Models. Mixed symbolic-geometric approaches encode both discrete relations (e.g., "left of," "inside") and numerical constraints (e.g., minimum distance) via polynomial encodings or bounding representations (Schultz et al., 2018, Chiatti et al., 2021).
• Neuro-Symbolic Integration. Recent ROS paradigms combine neural representation learning (scene parsing, attribute extraction) with explicit symbolic graph structures for relational and logical reasoning (Jahangard et al., 30 Oct 2025).
• 4D Discrete Space-Time Models. Systems such as ROSS posit space as a grid of unit-sized cells indexed by spatial $(x,y,z)$ and temporal $(t)$ coordinates, enabling fine-grained state reasoning (Hofford, 2014).

Key algebraic operations include converse, weak composition, and path-consistency constraint propagation, facilitating robust derivation of higher-order spatial facts from atomic relations.

2. Structured Spatial Reasoning Tasks and Benchmarks

Research in ROS leverages a growing suite of benchmarks targeting spatial inference beyond perception:

• RocketScience (Hoehing et al., 2 Sep 2025): A fully contrastive real-image/text benchmark for vision-LLMs (VLMs), probing the capacity to resolve challenging spatial relations such as "left of," "in front of," or absolute frame locations. Empirical analyses separate object localization ability from true spatial reasoning.
• SPaRC (Kaesberg et al., 22 May 2025): A suite of grid-based pathfinding puzzles with multi-layered arithmetic and geometric constraints (forbidden cells, shape-matching, region invariants), formalized as CSPs over grid graphs. Human solvers approach 98% accuracy; SOTA LLMs (o4-mini, GPT-4.1) are bottlenecked by failures in spatial logic, not perception.
• GeoReason-Bench (Li et al., 7 Jan 2026): A logic-driven dataset in remote sensing with 4,000 reasoning trajectories spanning geometric primitives and real-world expert knowledge, supporting detailed chain-of-thought (CoT) deduction and high-level spatial inference.

Such datasets are critical for exposing the cognitive and algorithmic gaps between symbol manipulation, statistical learning, and human-level spatial reasoning.

3. Learning and Inference Paradigms

ROS is realized through diverse computational paradigms, each targeting distinct reasoning capabilities and deployment contexts:

• Logical and Declarative Systems. Answer Set Programming Modulo Space-Time (ASPMT) (Schultz et al., 2018) enables declarative specification and constraint propagation for spatio-temporal reasoning, capable of integrating qualitative and polynomial (in)equality constraints for dynamic regions. Constraint propagation approaches in qualitative spatial networks or path-consistency algorithms underpin widespread spatial CSPs (Lee et al., 2022).
• Chain-of-Thought and Reinforcement Learning. Modern RS-VLMs (e.g., GeoReason (Li et al., 7 Jan 2026)) and LLM-based recommenders (e.g., next-POI ROS (Lv et al., 8 Jan 2026)) rely on multi-step CoT reasoning, often underpinned by two-stage training—initial supervised fine-tuning to internalize structured deduction, followed by reinforcement learning. Logical consistency rewards based on permutation-invariance of option ordering are essential for eliminating spurious CoT–decision decoupling.
• Neuro-Symbolic Approaches. Scene graphs with explicit 3D reasoning (Jahangard et al., 30 Oct 2025), structured probabilistic factorization (object proposal + 3D spatial relation scoring (Nejatishahidin et al., 2024)), and semantic feedback loops (symbolic consistency signals shaping neural network output channels (Lee et al., 2022)) realize scalable, interpretable ROS pipelines suitable for robotics.
• Complex Event and Language Understanding. Integrated ROS architectures, incorporating perception, cognitive semantics (e.g., IRL graph programs), and construction grammar, support robust generation and parsing of spatial language—even amidst perceptual errors and language omissions (Spranger et al., 2016).

4. Empirical Findings and Performance Analysis

Recent studies report critical performance dichotomies and limitations in both classical and contemporary models:

• Disentanglement of Perception and Spatial Reasoning. On benchmarks like RocketScience, SOTA VLMs are nearly optimal at object localization (above 95% accuracy) but consistently fail at the relational inference task (often at or below random chance without dedicated reasoning modules) (Hoehing et al., 2 Sep 2025). Injecting explicit CoT, reflection modules, or stepwise logical consistency rewards substantially boosts spatial reasoning.
• Constraint Satisfaction and Compositionality. In SPaRC, SOTA models frequently return invalid or incomplete solutions due to failure in enforcing global spatial constraints, logical fallacies in reasoning traces, or ignoring inter-rule interactions; humans consistently apply step-by-step checking and backtracking (Kaesberg et al., 22 May 2025).
• RL-Enhanced Consistency. In GeoReason, incorporation of option-permutation logical consistency rewards during RL raises Average Accuracy (AA) by +8–9 points above base SFT or vanilla policy learning, closing "wrong-reason–right-answer" failure modes and boosting interpretability (Li et al., 7 Jan 2026).

The following table profiles representative systems across paradigms:

System	Paradigm	Core ROS Mechanism	Performance Note
ASPMT	Symbolic/constraint	Qualitative-quantitative SAT checks	Scalable up to 40 objects/timesteps
GeoReason	Neuro-symbolic RL	CoT+Option-permutation logical reward	+12 OA gain vs open-source SOTA
SPaRC/o4-mini	LLM	Chain-step pathfinding + CSP constraints	15% solved vs. 98% human
Structured OVM	Modular probabilistic model	3D detection + learned 3D relation scoring	+20% accuracy vs. Detic or DINO VLM
JRDB-Reasoning	Scene graph + symbolic search	3D attribute-relation Prolog-style engine	+43–97% mAP vs. VLMs on ROS queries

5. Geographically Grounded and Generative ROS

Beyond static scene understanding, ROS is integrated into sequential decision and recommendation tasks:

• LLM-based Mobility Recommendation. ROS for next-POI leverages a Hierarchical Spatial Semantic ID encoding, three-stage Mobility CoT, and spatially guided reinforcement learning. This enables compact, locality-aware generative modeling for mobility prediction, yielding >10% HR@1 improvement over LLM-based and neural baselines and enhanced cross-city generalization (Lv et al., 8 Jan 2026).
• Spatially Explicit Reasoning in Remote Sensing. GeoReason demonstrates the necessity of synchronous reasoning chains and final decisions for reliable high-level spatial analytics (e.g., capacity estimation, zoning) using high-fidelity logic-trace datasets and CoT with RL alignment for RS-VLMs (Li et al., 7 Jan 2026).

6. Open Challenges and Future Directions

Key challenges and research avenues highlighted across the literature include:

• Robustness and Consistency. Handling contradictory or probabilistically uncertain knowledge bases; enforcing global path and region constraints; adaptive consistency verification (runtime monitoring) (Lee et al., 2022, Kaesberg et al., 22 May 2025).
• Scalable and Interpretable Reasoning. Efficient solution of large-scale spatial CSPs (incremental solvers, tractable subclasses) (Schultz et al., 2018, Lee et al., 2022); transparent, stepwise compositional inference in robotic and VLM systems.
• Generalization and Data Efficiency. Cross-domain and cross-city transfer in ROS-driven recommendation; curriculum and synthetic spatial data for end-to-end model training; leveraging symbolic priors to accelerate learning (Lv et al., 8 Jan 2026, Nejatishahidin et al., 2024).
• Higher-Order and Dynamic Reasoning. Extending logical consistency objectives to multi-step or interactive planning, integrating continuous geometric priors (vector maps, connected trajectories), and event-sequence understanding (Li et al., 7 Jan 2026, Hofford, 2014).
• Unified and Standardized Toolchains. There is an absence of comprehensive, benchmarked neuro-symbolic reasoning infrastructures and probabilistic qualitative spatial calculi for evaluation and deployment (Lee et al., 2022).

A plausible implication is that future ROS models will be explicitly multi-modal, integrating learned perception, symbolic graph structures, hierarchical semantic encodings, and consistency-enforcing training objectives, to achieve scalable, interpretable, and cognitively robust spatial reasoning across domains.