Researchers separate geographic memory from spatial generalization in language models
An arXiv paper argues that recalling trained location facts is not the same as predicting descriptions for places a model never saw, and tests continuous location representations as a fix.
In plain words
Large language models can link places to descriptions, but repeating a fact seen during training differs from describing an unfamiliar location. The authors add continuous location representations, supplied as spatial tokens, so geographic knowledge becomes organized across space. On new benchmarks for Los Angeles and London, this helped most at locations outside the training coverage.
What was announced
The paper distinguishes geospatial memory, meaning recovery of location-description associations observed during training, from spatial generalization, meaning prediction of descriptions for locations in unobserved regions. The authors argue that recovering an observed association does not by itself demonstrate generalization across geographic space. Their approach builds on a continuous spatial encoder, projecting location representations into a model's embedding space and supplying them as spatial tokens, which they frame as a latent spatial field linking descriptions through spatial dependencies.
Evaluation used two newly curated geo-text benchmarks covering Los Angeles and London. Geospatial captioning tested both memory and generalization by generating descriptions at observed locations and at locations beyond the spatial coverage of the training data. Across multiple LLM families, the continuous representations improved captioning at unobserved locations compared with textual coordinates, and the generated descriptions gave modest but consistent gains in a downstream grounding task that maps text to geographic space. Layer-wise probing and representation clustering offered further evidence consistent with spatially organized geographic representations. The authors present this as support for evaluating generalization separately from memory.
What it does not show
- The evidence is an abstract only; full experimental details, model sizes and benchmark construction are not available in the supplied passages.
- Improvements in downstream grounding are described as modest, and no independent replication of the results is reported.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.