Spatial Reasoning via Modality Switching Between Language and Symbolic Representations

Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

PDF
Code

Shreya Rajpal, Tanawan Premsri, and Parisa Kordjamshidi.

We introduce an adaptive modality-switching framework for complex multi-hop spatial stories. The approach selects natural-language or symbolic representations using trustworthiness and complexity signals, improving performance on complex spatial reasoning by 42%.