COSIT 2026: Exploring Spatial Reasoning and Perspective Transformation Capabilities in Language Models

Title: COSIT 2026: Exploring Spatial Reasoning and Perspective Transformation Capabilities in Language Models Core Abstract: This study focuses on the performance of language models in spatial reasoning and perspective transformation tasks. By constructing a dedicated dataset and experimental framework, it evaluates the capability boundaries of mainstream models, providing an empirical foundation for spatial intelligence and embodied AI. The study has open-sourced the complete dataset and experimental code to support reproduction and expansion.

Research Background and Motivation

Spatial reasoning is a core component of human intelligence, permeating daily life (navigation, object manipulation, etc.). Although LLMs perform well in NLP tasks, their spatial reasoning capabilities have not been fully explored. Perspective transformation is a key spatial reasoning capability that requires an agent to understand a scene from different perspectives, which is crucial for applications such as collaborative robots and autonomous driving. This study fills the gap by systematically evaluating the performance of LLMs in spatial reasoning and perspective transformation, and the open-sourced data and code provide a foundation for subsequent research.

Research Methods and Dataset Construction

Data Structure Design

The dataset uses JSON Lines format to store training/test sets. The data_generator folder contains generation scripts that support understanding logic, reproducing data, generating variants, and expanding scenarios.

Experimental Framework

• Main Experiment Script (main.py): Executes fine-tuning and testing, supports command-line parameters (check via python3 main.py --help).
• Generalization Experiment Script (generalisation.py): Fine-tunes DeBERTa-v3-large on simple instances and tests generalization ability on OOD instances (check via python3 generalisation.py --help).

Experimental Results and Model Predictions

Prediction Result Storage

The predictions folder stores fine-tuned model predictions grouped by random seeds, facilitating tracking differences, statistical tests, and stability analysis.

Gemini Prediction Results

predictions/gemini_pred stores Gemini model predictions, serving as a baseline to compare performance gaps between general models and fine-tuned models, as well as the characteristics of different architectures and the help of pre-trained knowledge.

Research Contributions and Significance

Theoretical Contributions

Provides empirical data for spatial information theory, revealing the capability boundaries, failure modes, and improvement directions of LLM spatial reasoning.

Practical Value

• Benchmark testing tool: Evaluate spatial reasoning capabilities of new models;
• Fine-tuning guide: Demonstrate optimization methods for spatial reasoning tasks;
• Data generation template: A starting point for building larger-scale datasets.

Open-Source Contributions

Provides complete datasets, reproducible code, detailed documentation, and standardized citation formats.

Technical Details and Usage Recommendations

Environment Configuration

Dependency inference: Python3.x, PyTorch/Transformers, jsonlines, numpy, pandas, etc. It is recommended to check the script import statements to install dependencies.

Experiment Reproduction Process

1. Data exploration: View samples in the data folder;
2. Data generation: Run the generator script;
3. Baseline experiment: Run the main experiment using main.py;
4. Generalization test: Evaluate using generalisation.py;
5. Result analysis: Compare predictions output with paper metrics.

Extended Research Directions

Multimodal expansion, testing of larger-scale models, cross-language evaluation, practical applications (robot navigation, etc.).

Limitations and Future Work

Current Limitations

1. Limited data scale, scene types need to be expanded;
2. Narrow model scope (mainly DeBERTa and Gemini);
3. Single task type (focus on perspective transformation).

Future Directions

1. Expand the dataset;
2. Include more mainstream models for comparison;
3. In-depth analysis of model mechanisms;
4. Apply to practical spatial intelligence systems.

Conclusion

Spatial reasoning is a key piece of the general AI puzzle. Through rigorous dataset construction and experimental evaluation, this study provides empirical data on the spatial reasoning capabilities of LLMs. The open-source release promotes academic reproduction and industrial applications, and we look forward to inspiring more spatial intelligence research and advancing AI's ability to understand spatial relationships in the physical world.