acceptodds
Under review as a conference paper at ICLR 2027

From Geographic Memory to Spatial Generalization in Large Language Models

Abstract

Large language models (LLMs) can associate places with geographic descriptions, but recovering an observed association does not establish an ability to generalize across geographic space. We distinguish geospatial memory, the recovery of location–description associations observed during training, from spatial generalization, the prediction of descriptions for locations in unobserved regions. We investigate whether continuous location representations improve this generalization and encourage geographic knowledge to become spatially organized within an LLM. We frame this organization as a latent spatial field: a learned mapping from locations to descriptions connected through spatial dependencies. Building on a continuous spatial encoder, we project location representations into the model's embedding space and supply them as spatial tokens. We evaluate the resulting models on two newly curated geo-text benchmarks for Los Angeles and London using two complementary assessments. Geospatial captioning tests both memory and spatial generalization by generating descriptions at observed locations and at locations beyond the spatial coverage of the training observations. A separate geospatial grounding model tests whether generated location–description pairs provide useful additional supervision for mapping text to geographic space. Across multiple LLM families, continuous location representations improve captioning at unobserved locations relative to textual coordinates, while their generated descriptions yield modest but consistent improvements in downstream grounding. Layer-wise probing and representation clustering provide additional evidence consistent with spatially organized geographic representations. Together, these findings show the value of evaluating spatial generalization separately from memory and of incorporating continuous spatial representations into language models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM, Geospatial Reasoning, Spatial Representation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_d8VKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work differentiates geographic memory and spatial generalization in LLMs, and frames geographic knowledge as a latent spatial field. The authors integrate continuous location embeddings as spatial tokens into LLMs, comparing with textual, grid and geohash representations on Airbnb-derived Los Angeles and London benchmarks. Evaluated via geospatial captioning and extrinsic geospatial grounding, continuous representations outperform alternatives for spatial generalization, even though textual coordinates excel at memorization.

Critical strengths

The paper clearly distinguishes geospatial memory from spatial generalization through observed and spatially held-out locations.

Critical weaknesses
  1. The experiments are limited to Airbnb descriptions from Los Angeles and London, which may introduce domain-specific biases. It remains unclear whether the findings generalize to other cities, geographic attributes, or data sources.
  2. The proposed method mainly integrates an existing Geo2Vec encoder into LLMs through a learned projection, with limited architectural innovations.
  3. The paper defines spatial reasoning primarily as predicting descriptions in unobserved regions, but this may reflect spatial interpolation rather than genuine geographic reasoning. The tasks requiring explicit spatial relationships or multi-step inference would better support the 'reasoning' claim.
  4. The improvements from continuous spatial representations are relatively modest, particularly in downstream grounding.
  5. The probing and clustering results do not conclusively demonstrate that LLMs learn a latent spatial field, as the observed spatial structure may largely originate from the pretrained continuous location encoder rather than newly acquired geographic knowledge.
Questions for authors
  1. Could the authors conduct an additional experiment comparing the spatial structure of Geo2Vec embeddings with LLM hidden representations before and after fine-tuning?
  2. Could the authors provide statistical significance tests for the improvements of LATENT over TEXT and HYBRID, particularly in downstream grounding?

I would be willing to raise my score if the authors can address my concerns.

2 comments

We thank the reviewer for the detailed comments. We respond to each point below.


Q1 and W5 (does the spatial structure come from the encoder?)

The encoder is not pretrained. The location encoder and projection are trained from scratch together with the LoRA adapters. Before training, the encoder carries only the geometry of the coordinates and no geographic knowledge.

Additional experiment. As requested, we compared the spatial structure of the location encoding with the LLM hidden representations; the result is in Figure 13 of the updated manuscript. The encoding alone preserves only spatial proximity. After the LLM learns from the data, the clusters reflect the data. For example, the coastal areas around Santa Monica and Long Beach are grouped together, following the coastline. Geometry alone cannot produce this grouping.


Q2 and W4 (modest improvements, significance tests)

Captioning. Figure 7 shows the comparison as the distance growth, with 95% bootstrap confidence intervals. The gap between LATENT and TEXT is significant in 20 out of 30 distance ranges.

Additional statistical test. We added a per-metric significance test that combines the three LLM families and both cities (two-sided stratified t-test with Holm correction; full results can be found in the updated version of submission at Table 7 in the appendix). To sum up, LATENT is significantly better than TEXT on Hit Rate@1% (+0.43 points, p<0.001), Hit Rate@5% (+0.37 points, p=0.006) and JSD (−0.007, p<0.001). Moreover, LATENT is also the only augmentation that significantly improves on the grounding model trained without augmentation. TEXT shows no significant change.

How we read this. Each single gain is modest, but it reaches significance on three of five metrics against TEXT, and the direction is consistent across model families and cities. Grounding is also an indirect test, so the room for improvement is small. We will revise the wording in Section 6.5 to match these tests.

On the size of the improvement. We agree that the gains are modest. Our contribution is to formulate and evaluate a concrete approach for incorporating continuous spatial representations into LLM geospatial reasoning, an initial step toward a difficult, broader goal. The grounding results suggest that better spatial representations alone are insufficient, helping identify where further methodological advances are needed. Beyond the immediate performance gains, the work provides an explicit formulation, an evaluated method, and an empirical reference point that subsequent research can build on to improve spatial reasoning and grounding. We will add the discussion into the paper.


W1 (data limited to Airbnb, Los Angeles and London)

Why this data. To help an LLM form geospatial awareness, we need a large number of human-written descriptions of locations that are coupled with coordinates. To the best of our knowledge, Airbnb listings are one of the few openly available, high-quality sources that provide this in detail at city scale, and no prior work has used them in this way. Other geo-tagged text, such as reviews or photo captions, mostly describes the business or the photo and not the location itself. We therefore see the new use of this data, together with the released training data and benchmarks, as one of the contributions of the paper. Together the two metropolitan areas provide about 446,000 location–description pairs, which we believe is the best coverage currently available for this purpose.

Evidence of generality within the study. The two cities are in different countries and have different urban layouts, and we train on each independently with three LLM families. The consistency experiment performance suggests that our finding is not specific to one city or one model.

W2 (limited architectural innovation)

Why the architecture is simple. The current multimodal LLM pipeline, an encoder followed by a learned projection, has been verified to be effective, so we see no reason to make it more complex.

Where the contribution lies. Our contribution is not a complex architecture. It lies in the framing of how an LLM can build geospatial awareness through memory and generalization, and how the representation of location affects this ability, which to our knowledge has not been proposed before. The open datasets, which support training as well as benchmarking, are also contributions to the community. A simpler and robust representation module is also more likely to be adopted in future LLM systems.


W3 (reasoning vs. interpolation)

On the term. We agree with the reviewer that “generalization” is the more conservative term, and we use it in our title and abstract. In the paper we also state that the reasoning we refer to is this generalization capacity. We keep “geospatial reasoning” as a defined term because we believe reasoning over geographic space differs from deductive reasoning, which requires explicit relationship inference. The revision will state this definition explicitly at first use. Following the reasons:

Why geospatial reasoning should be different. Deductive reasoning derives a conclusion from relations stated in the input. Geographic data does not have this form, since the geospatial space is continuous, while relationships saved in natural language are discrete. Also, most geographic dependencies are implicit, and it is impossible to state all the relevant world knowledge in the input. We therefore study geospatial reasoning in a latent, implicit form, where the LLM infers from its internal knowledge. We see this setting as complementary to relational reasoning: those settings assume the relations are given, while ours addresses the case where they are not.

Why this is not interpolation. Geographic knowledge expressed in natural language differs from numeric values. Spatial interpolation averages the numeric values of nearby observations, while a text description has no direct analog of this operation. The closest counterpart for text data is to copy descriptions from the nearest observed locations, which is our k-NN baseline. In the experiments, the tuned LLMs with every representation outperform k-NN in held-out regions (Table 2), and more importantly, this gap grows as test locations lie farther from the training data (Fig. 7). The model is therefore doing more than reusing nearby text, which is why we describe it as reasoning.

Additional evidence. Appendix A.9 asks each model for the distance from a location to LAX. The seven locations were manually selected from regions that contain no training observations. For example, points 0, 4 and 6 are at a canal, on a mountain and in a lake, where the closest POI is kilometers away. LATENT’s estimates rise with the true distance, from 2.4 miles to 25 miles. This qualitative case study provides evidence that LLM can generalize to distant and no data regions and support our claim.

Reviewer_AVEnOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies geospatial awareness of LLMs and distinguishes between two abilities: memory and generalization. The authors develop this characterization through the perspective of a latent spatial field. The authors evaluate 3 LLM families on two tasks: geospatial captioning and geospatial grounding, and find that continuous location representations can improve the generalization of LLMs.

Critical strengths
  • The writing is generally clear, though the introduction is overly long and could be more concise.
  • The introduced location–description dataset constructed from Airbnb listings is useful for future work in this field, and the construction details are well documented.
Critical weaknesses
  • My main concern is that the contribution doesn’t meet the bar of ICLR. The characterization of memory and generalization is not new: it amounts to comparing performance on training locations with performance on spatially held-out locations, which is a standard practice in geospatial ML[1-3]. Also, the finding that continuous location representations help LLM generalization is expected, since continuous encoders are designed to give nearby locations similar representations, which naturally helps prediction at nearby unseen locations. Overall, this paper offers limited new insights.
  • My second concern comes from the soundness of experiments. The held-out regions are small (grid cells of around 0.5 km in Los Angeles and 1.2 km in London) and close to the training data, so I think the current evaluation is more like short-range interpolation than meaningful generalization to truly uncovered regions.

[1] Roberts, David R., et al. "Cross‐validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure." Ecography 40.8 (2017): 913-929.

[2] Ploton, Pierre, et al. "Spatial validation reveals poor predictive performance of large-scale ecological mapping models." Nature communications 11.1 (2020): 4540.

[3] Klemmer, Konstantin, et al. "Satclip: Global, general-purpose location embeddings with satellite imagery." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 4. 2025.

Questions for authors
  1. Have the authors run experiments on truly uncovered regions, e.g., by holding out much larger areas (such as entire neighborhoods) or by testing on a city not seen during training? If so, do continuous location representations still outperform textual coordinates in that setting?
2 comments

We thank the reviewer for the comments. We respond to each point below.


W1: Contribution and relationship to prior work

On the contribution. We acknowledge that spatial validation and continuous location representations are established. However, our contribution is not the spatial holdout itself or showing that a continuous encoder is better. Our contribution concerns how these representations affect the acquisition, organization and generalization of geographic knowledge within LLMs. To the best of our knowledge, this is the first systematic framework for this question.

How our setting differs. Standard spatial validation tests prediction from features to a target. Our setting is prediction from the coordinate alone, which is harder: no information about the held-out location reaches the model, so it must infer from internalized knowledge. Unlike the domain-specific models in the cited works, an LLM is expected to hold such world knowledge.

Relation to SatCLIP. SatCLIP pretrains its representation with a random 90%/10% split and no spatial holdout; its spatial holdouts apply only to downstream prediction after global image–location pretraining. Our holdout happens during coordinate–description learning, which corresponds to SatCLIP’s pretraining stage. We therefore expect the LLM to fill the gap from its own knowledge, whereas SatCLIP is not expected to do so: it relies on imagery of those regions observed during pretraining.

Beyond the expected smoothness. The inductive bias of continuous encoders is expected, but to our knowledge no prior work has systematically tested how it affects the memory and generalization of geographic knowledge in LLMs, and its effect on LLMs is not guaranteed. We also find that strong recovery of observed associations does not imply the strongest generalization: TEXT is ahead of LATENT on memory, and the order reverses at held-out locations. Together with the representational analyses, these are the insights beyond previous papers.

In short. Three things are new in this paper. First, we test how location representations affect the way an LLM acquires, organizes and generalizes geographic knowledge. Second, we find a trade-off between memory and generalization and provide an analysis that helps explain it. Third, we provide the datasets and benchmarks that make this test possible.


W2/Q1: Uncovered regions and evaluation scale

Evidence at locations far from training data. Please refer to the qualitative analysis in Appendix A.9 (Figure 11 and Table 6). We deliberately sampled seven query points from data-sparse held-out regions that contain no training observations. For example, points 0, 4 and 6 are at a canal, on a mountain and in a lake, where the closest POI is kilometers away. LATENT gives the closest estimate of the distance to LAX at points 0, 1 and 6, and its estimates rise with the true distance across the seven points. This is a qualitative case study, and it shows the model responding to location where it has no data. The distance-stratified analysis (Appendix A.5, Figure 7) supports this quantitatively. we group the held-out locations by their distance to the nearest training observation, and we compare the methods within each distance range. The advantage of the continuous representation grows with this distance. LATENT performs best in the farthest distance range in five of six model–city settings. This supports our claim regarding generalization.

What the paper tests. Our scope is to test the geospatial memory and generalization capabilities of LLMs using a representation of space itself, rather than through textual association. We want to understand whether an LLM can memorize space as part of its world knowledge. Knowledge cannot come from nowhere. Therefore, in this paper, we study data sparsity in a controlled form: local gaps inside a region that the model has observed. This is a more realistic setting. In a city, descriptions cluster around populated blocks, and the places without descriptions are the gaps between them.

On unseen cities. Testing on an unseen city is outside the scope of this research, and we do not expect the model to generalize to one. Knowledge about a city, such as its landmarks and neighborhoods, must come from observations of that city. What transfers across cities is the method: we train independently on Los Angeles and London with three LLM families and observe consistent results.

Thank you for the detailed response. However, I still think the contribution is not significant enough for ICLR. In addition, Table 6 in the appendix shows that LATENT gives the closest prediction in only 3 out of 7 cases, while another baseline (GRID) is also closest in 3 cases and seems to have a better average prediction across all 7 cases, so this case study is not convincing enough to support the generalization claim. Therefore, I will keep my score unchanged.

Reviewer_XktROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether Large Language Models (LLMs) memorize geospatial knowledge via surface-level text associations or if they can build an internal "latent spatial field" for spatial generalization. To address this, the authors propose a "Spatially Explicit LLM" that incorporates a continuous spatial encoder to inject location representations as continuous tokens (LATENT/HYBRID), replacing or augmenting discrete textual coordinates. The authors curate a dataset from Airbnb listings in Los Angeles and London, utilizing a spatial block-split strategy to separate "seen" (memory) and "unseen" (generalization) regions. The models are evaluated on geospatial captioning and grounding tasks, complemented by layer-wise probing and hidden-state clustering to visualize the learned spatial field.

Critical strengths

Comprehensive Experimental Setup for Coordinate Representation: The paper features a well-designed, multi-faceted experimental framework (including intrinsic captioning, extrinsic grounding, layer-wise probing, and clustering). It effectively demonstrates that, under the strict constraint of only using 2D coordinate inputs, continuous spatial embeddings have a distinct advantage over textual or discrete 2D coordinates. It convincingly shows that it is extremely difficult for LLMs to learn or activate valid geographic information solely from raw 2D coordinate text.

Compelling Qualitative Visualizations: The spatial clustering of hidden states (Figures 9 and 10) provides intuitive and strong qualitative evidence. It effectively demonstrates that continuous spatial encodings help the model form smoother, more geographically coherent spatial boundaries compared to the fragmented representations produced by discrete or pure textual coordinates.

Critical weaknesses

Lack of Practical Baselines and Ignored Semantic Context (Address/POI): The most critical flaw of this paper lies in its restricted experimental setup, which artificially isolates the LLM from its greatest strength: text semantics. In real-world applications, coordinates are rarely fed to LLMs as raw numbers or continuous vectors alone; they are reverse-geocoded into rich textual contexts (e.g., Address, nearby POIs). The paper fails to compare the computationally expensive LATENT method (which requires LoRA fine-tuning and training an encoder from scratch) against a highly practical, zero-shot/RAG-style baseline: TEXT (Coordinate + Address + POI). Without this comparison, it is impossible to know whether the geometric benefits of the "latent spatial field" are merely marginal when rich semantic spatial contexts are provided, thus making the proposed method's real-world utility highly questionable.

Misaligned Claims on "Reasoning": The paper explicitly claims to evaluate "geospatial reasoning." However, successfully generating a caption for a held-out location based on training data from adjacent grid cells (often just hundreds of meters away) is fundamentally "spatial interpolation" or "spatial smoothing." It does not constitute logical, topological, or deductive "reasoning" (e.g., inferring spatial relations). The terminology overclaims the actual capability demonstrated by the experiments.

Overclaimed Impact on "Data Sparsity": The authors motivate their work by highlighting the macro-level sparsity of real-world geographic data (e.g., urban vs. rural areas). Yet, the experiments are exclusively conducted on micro-scale intra-city grids within two of the most heavily documented global metropolises (Los Angeles and London). Bridging a ~500m to 1km gap in LA does not substantiate the claim of solving or addressing realistic geospatial data sparsity.

Missing Citations and Comparisons: The paper overlooks important previous works that also explore spatial representations and knowledge extraction in LLMs. Specifically, the paper misses discussions and comparisons with "Geolocation representation from large language models are generic enhancers for spatio-temporal learning" (AAAI 2025). Furthermore, while "GeoLLM: Extracting geospatial knowledge from large language models" is cited, the paper fails to structurally compare its baselines against the rich knowledge extraction paradigms established in such works. Crucially, both of these prior studies have already recognized and documented that raw 2D coordinates, when treated as plain text, provide almost no meaningful geographic utility for LLMs. The current paper somewhat frames the limitation of textual coordinates as a primary motivation, without properly attributing this established consensus to the existing literature.

Questions for authors

Practical Baseline: How would the LATENT and HYBRID representations perform if the TEXT baseline was augmented with reverse-geocoded spatial text (e.g., exact address, top 5 nearby POIs)? Does the rich semantic context of such text overshadow the geometric benefits provided by your continuous spatial embeddings? Can you provide empirical results for this baseline?

Reasoning vs. Interpolation: Can you justify the use of the term "reasoning" for your task? Given that the model is predicting descriptions for held-out grids based on nearby seen grids, how does this differ from traditional spatial interpolation?

Macro-level Sparsity: Your motivation relies heavily on data sparsity, but LA and London are incredibly data-dense. Do you have evidence that your method can generalize across macro-level sparsity (e.g., training on a city and testing on rural areas, or zero-shot transfer to an entirely unseen city)?

5 comments

We thank the reviewer for the detailed comments. We respond to each point below.


W1 (semantic context)

On the scope. We agree with the reviewer that for retrieval, a RAG-style method, or querying an LLM with reverse-geocoded addresses and POIs, is more efficient and effective than querying by a spatial representation (e.g. coordinate, geohash) directly. However, the scope of this paper is not to propose a retrieval method for geospatial information.

What the paper tests. Our scope is to test the geospatial memory and generalization capabilities of LLMs using a representation of space itself, rather than through textual association. The question is whether geographic knowledge can be stored in the model and extended to locations without observations. In the longer term, we want to understand whether an LLM can memorize space at a global scale as part of its world knowledge. The two cities studied here are a first, controlled step toward that question.

Why a retrieval baseline answers a different question. A prompt with address and POI text supplies the knowledge from an external database at inference time. Its performance reflects the coverage of that database, not what the model has internalized, which is what we test.

Why retrieval alone cannot solve geospatial reasoning. A retrieval set is discrete and finite. Geographic space is continuous. Relations such as “close to” can be ambiguous and hold for infinitely many locations, so it cannot all be expressed as explicit records.


Q1 (practical baseline)

Additional experiment. We added a baseline that follows the prompt design of GeoLLM. The location is represented by its coordinate, street address and the four nearest POIs (e.g., “(lat,lon), 1201 S Figueroa St, Los Angeles; Los Angeles Convention Center, Crypto Arena, 7-Eleven, Team LA”), and the model is fine-tuned under the same protocol. The held-out captioning results with Qwen3.5 are:

RepresentationLos Angeles BERTScoreLos Angeles ROUGE-LLondon BERTScoreLondon ROUGE-L
GeoLLM0.53300.30880.52070.2959
TEXT0.53450.32430.55800.3377
LATENT0.54320.32120.56070.3455
k-NN0.52130.29310.53060.2980

The address and POI representation does not generalize well. It is below LATENT on every metric. The scope of this research is to find a universal way to represent space for an LLM, so that the LLM can better perceive space. Address and POI data cannot serve as a general representation at larger scales


W2 / Q2 (reasoning vs. interpolation)

On the term. We agree with the reviewer that “generalization” is the more conservative term, and our title and abstract use it. In the paper we also state that the reasoning we refer to is this generalization capacity. We keep “geospatial reasoning” as a defined term because we believe reasoning over geographic space differs from mathematical or deductive reasoning, the revision will state this definition explicitly at first use. Following the reasons:

Why geospatial reasoning is different. Deductive spatial reasoning derives a conclusion from relations stated in the input. Geographic data does not have this form. Geographic space is continuous while natural language is discrete, so the relations among locations cannot all be written as statements. Most geographic dependencies are also implicit, and it is impossible to state all the relevant evidence in the input. We therefore study geospatial reasoning in a latent, implicit form, where the LLM infers from its internal knowledge.

Why this is not interpolation. Geographic knowledge expressed in natural language differs from numeric values. Spatial interpolation averages the numeric values of nearby observations, while a text description has no direct analogue of this operation. The closest counterpart for text data is to copy descriptions from the nearest observed locations, which is our k-NN baseline. In the experiments, the fine-tuned LLMs outperform k-NN in held-out regions (Table 2), and more importantly this gap grows as test locations lie farther from the training data (Fig. 7). The model is therefore doing more than reusing nearby text, which is why we describe it as reasoning.

Internalized knowledge. Geospatial reasoning, as we define it, is this inference from internalized knowledge. We see this setting as complementary to topological and deductive reasoning. Our experiments are designed for our question: the same captioning task at observed and held-out locations separates memory from reasoning.

W3 / Q3 (data sparsity)

Evidence at locations far from training data. Please refer to the qualitative analysis in Appendix A.9 (Figure 11 and Table 6). We deliberately sampled seven query points from data-sparse held-out regions that contain no training observations. For example, points 0, 4 and 6 are at a canal, on a mountain and in a lake, where the closest POI is kilometers away. LATENT gives the closest estimate of the distance to LAX at points 0, 1 and 6, and its estimates rise with the true distance across the seven points. This is a qualitative case study, and it shows the model responding to location where it has no data. The distance-stratified analysis (Appendix A.5, Figure 7) supports this quantitatively: LATENT performs best in the farthest distance bin in five of six model–city settings.

On unseen cities. Testing on an unseen city is outside the scope of this research, and we do not expect the model to generalize to one. We want to understand whether an LLM can memorize space as part of its world knowledge. Knowledge cannot come from nowhere. Knowledge about a city, such as its landmarks and neighborhoods, must come from observations of that city. What transfers across cities is the method: we train independently on Los Angeles and London with three LLM families and observe consistent results.


W4 (missing citations)

On the citation and attribution. We thank the reviewer for pointing this out. We will add LLMGeovec (He et al., AAAI 2025) to the related work, expand our discussion of GeoLLM, and add a GeoLLM-style prompt as a baseline (see our response to Q1). We will also attribute to these works the finding that raw coordinates are a weak interface to a pretrained LLM. Our study adds a new finding on top of it: raw coordinates are weaker in generalization, but they perform well in memorization once the knowledge is covered during training. This refines the earlier finding and does not repeat it.

How our scope differs. Both works extract knowledge that the pretrained LLM already holds, and they reach it through text: the prompt names the address and nearby places, and the LLM connects those names to the requested information. Our study deliberately avoids this kind of textual association, because we want the LLM to hold knowledge of space itself. A structural comparison with the knowledge-extraction paradigm would therefore test a different question.

Additional experiment: memory performance of the GeoLLM (follow-up to Q1)


Following our response to Q1, we also tested the GeoLLM baseline on the memory task, at observed locations. The model is fine-tuned under the same protocol. The captioning results at observed locations with Qwen3.5 are:

RepresentationLos Angeles BERTScoreLos Angeles ROUGE-LLondon BERTScoreLondon ROUGE-L
TEXT0.56230.35850.58490.3821
GeoLLM0.56870.37750.57570.3709
LATENT0.55360.34010.57710.3711

In Los Angeles, GeoLLM performs best on the memory test and outperforms the textual coordinate. The address and POI names give the model additional text for each observed location, so it memorizes these locations more accurately.

The benefit is not consistent across cities. In London, GeoLLM is below TEXT and no better than LATENT. Its benefit depends on the address and POI data of the city. This supports our claim that address and POI augmentation is not a universal way to represent space. It does not let an LLM perceive space directly.

Additional experiments further clarify our contribution. As shown in our response to Q1, GeoLLM is below LATENT on every metric at held-out locations. Taken together, these experiments further support our claim. A representation built on textual association helps the model memorize observed locations, but this knowledge does not extend well to locations without observations. A continuous representation of space itself generalizes better.

To the best of our knowledge, this paper is the first to clearly propose and test the difference between geospatial memory and generalization when an LLM perceives space directly. We expect it to be of broad interest to both the LLM and geospatial communities.

We hope this response makes the scope of the paper clearer: it studies how an LLM perceives space directly, not how to retrieve geographic information. We would be grateful if you could reconsider your evaluation from this perspective, and we are happy to answer any further questions. Thank you again for your time and thoughtful feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.