Learning to Draw ASCII Improves Spatial Reasoning in Language Models

ACL ARR 2026 January Submission4502 Authors

Keywords: Spatial Reasoning, Interpretability, Large Language Models, Reasoning, ASCII Representation

Abstract: Human spatial cognition involves both perceiving spatial configurations and reconstructing them through sketching or mental imagery. Although perceiving is often intuitive, reconstruction is substantially more demanding. To facilitate this process, humans frequently rely on sketching as a "read-write loop" to visualize ideas and anchor their reasoning when solving complex spatial problems. In this work, we investigate whether Large Language Models (LLMs) exhibit a similar dependency when processing spatial topology. We introduce Text2Space, a dataset that pairs natural language descriptions with ground-truth ASCII representations and QA pairs, using ASCII as a diagnostic probe to decouple representation failures from reasoning failures. Our evaluation reveals a pronounced "Read-Write Asymmetry": LLMs interpret ASCII representations effectively but struggle to reconstruct them, and this generation bottleneck propagates to the subsequent reasoning phase, producing flawed intermediate representations that distort subsequent reasoning. Training models on Text$\leftrightarrow$ASCII alignment significantly improves performance on text-only spatial reasoning tasks, even without generating ASCII at inference time. Our findings show that enabling LLMs to "learn to draw" not only allows them to better leverage sketching as a reasoning strategy but also fosters deeper internalization of spatial understanding.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: Interpretability and Analysis of Models for NLP, Language Modeling, Question Answering

Contribution Types: Model analysis & interpretability, Data resources

Languages Studied: English

Submission Number: 4502