More than Correlation: Do Large Language Models Learn Causal Representations of Space?
This addresses a criticism in AI research about whether LLMs' spatial understanding is causal, which is important for improving their reliability in geospatial applications.
The study investigated whether large language models (LLMs) learn causal spatial representations rather than just correlations, finding through intervention experiments that these representations influence performance on next-word prediction and geospatial tasks.
Recent work found high mutual information between the learned representations of large language models (LLMs) and the geospatial property of its input, hinting an emergent internal model of space. However, whether this internal space model has any causal effects on the LLMs' behaviors was not answered by that work, led to criticism of these findings as mere statistical correlation. Our study focused on uncovering the causality of the spatial representations in LLMs. In particular, we discovered the potential spatial representations in DeBERTa, GPT-Neo using representational similarity analysis and linear and non-linear probing. Our casual intervention experiments showed that the spatial representations influenced the model's performance on next word prediction and a downstream task that relies on geospatial information. Our experiments suggested that the LLMs learn and use an internal model of space in solving geospatial related tasks.