acceptodds
Under review as a conference paper at ICLR 2027

Learning Administrative Boundary-Aware Location Encoders via Reverse Geocoding

Abstract

Location encoders learn compact representations of latitude–longitude coordinates for downstream geospatial prediction tasks. Existing approaches are pretrained using continuous or naturally occurring supervision, such as imagery, climate variables, or species distributions, and therefore do not explicitly capture administrative boundaries or place names, despite their strong influence on many real-world geospatial phenomena. We introduce, to our knowledge, the first method for learning administrative boundary-aware location encoders, using reverse-geocoded text as a self-supervised training signal. Our approach can be applied to existing location encoder architectures and consistently improves their downstream performance. We further introduce RevGeoRQ, a custom architecture tailored to reverse-geocoding supervision, which yields additional gains and matches or exceeds contemporary location encoders on several downstream tasks. Finally, the use of human-interpretable reverse-geocoded strings and residual-quantized semantic IDs provides a more interpretable representation of geographic location than existing approaches.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
geospatial machine learning, location encoders, self-supervised learning, reverse geocoding, geospatial representation learning, location embeddings, geographic representation learning, multimodal learning, geographic priors

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_meqgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a self-supervised framework for learning boundary-aware location embeddings. Unlike existing location encoders, such as GeoCLIP, which typically learn geographic representations through visual supervision, the proposed approach leverages reverse-geocoded text to incorporate administrative and geographic information. Experimental results suggest that the learned embeddings capture information complementary to existing geographic representations, and combining them can improve performance on several downstream tasks. However, the method has limited technical novelty, and the paper lacks an in-depth analysis of the potential biases introduced by administrative boundary-aware supervision.

Critical strengths
  • The proposed self-supervised framework derives supervision signals from reverse-geocoded text, incorporating administrative boundaries and geographic place names into location representations. This provides a complementary source of information to the visual or environmental supervision commonly used in existing location encoders.

  • The experiments demonstrate that combining the proposed embeddings with existing geographic representations can improve performance on several downstream tasks. The evaluation covers multiple tasks and integrates the proposed approach with several existing location encoders, providing evidence of its potential applicability and complementarity.

Critical weaknesses
  • The proposed framework primarily integrates existing techniques, including reverse geocoding, pretrained text embeddings, and RQ-VAE. The reverse-geocoded text consists of only five administrative attributes (city, admin1, admin2, country, and continent), which are essentially categorical geographic identifiers rather than rich natural-language descriptions. It remains unclear whether encoding these simple attributes using a pretrained Qwen3-Embedding model offers meaningful advantages over conventional categorical embeddings or simpler alignment methods.

  • Administrative supervision may bias the learned representations toward boundary-aware geographic characteristics while weakening their ability to capture other important spatial properties, such as environmental similarities. This could introduce undesirable biases and reduce representation diversity and generalizability. The authors should provide a more in-depth analysis of these trade-offs rather than relying primarily on aggregate performance improvements.

  • The proposed approach relies exclusively on simple administrative attributes without exploring richer semantic information, such as climatic or socioeconomic characteristics. This potentially limits the diversity and expressiveness of the learned representations. The paper does not sufficiently justify why such a restricted set of geographic attributes is appropriate for general-purpose location representation learning.

  • Reverse-geocoded information is central to the proposed framework, yet the paper lacks ablation studies examining the contributions of different administrative attributes (e.g., city, region, country, and continent). It remains unclear whether coarse-grained information, such as country labels alone, could achieve comparable performance. Such experiments are important for establishing the effectiveness and necessity of the proposed supervision strategy.

Questions for authors
  • What motivates the choice of RQ-VAE and Qwen3-Embedding? Have the authors considered other alternatives, such as learnable categorical embeddings or direct embedding alignment? How do these alternatives compare in terms of effectiveness and computational efficiency?

  • How accurately do the learned embeddings capture actual administrative boundaries? Which types of downstream tasks or geographic scenarios benefit most from boundary-aware supervision, and in which cases might it introduce undesirable biases or degrade performance?

  • Why does the proposed framework rely exclusively on a relatively simple reverse-geocoding strategy? Have the authors considered extracting richer geographic semantics to improve representation diversity and generalizability?

  • How does the performance change when different levels of administrative information are used for supervision (e.g., country only, country plus region, or the full administrative hierarchy)? Can the authors demonstrate that fine-grained administrative information provides meaningful improvements over coarse-grained labels?

4 comments
AuthorsOpenReview ↗

Rebuttal response to Reviewer meqg (1/n)

We would like to thank Reviewer meqg for their review. Please find our responses and improvements below:

  • W1: "It remains unclear whether encoding these simple attributes using a pretrained Qwen3-Embedding model offers meaningful advantages over conventional categorical embeddings or simpler alignment methods."
    • Response: We agree that an experiment with categorical embeddings makes the study more thorough, and would like to thank the reviewer for the suggestion. We’re sharing the results below, and will include this in the completed manuscript.
      • Result: the new hierarchical-ID baseline takes a lat-long input, reverse geocodes it, and produces a tree-like hierarchical ID. This proves to be a strong baseline with an average .669R2.669 R^2 on regression tasks and .676.676 acc. on classification tasks. However, it significantly trails all the variants trained with our method. For illustration, our Custom RevGeoRQ model achieves an average .835R2.835 R^2 on regression tasks and .841.841 acc. on classification tasks. Full results:

        Modelair tempelevationpopulationcali housingavg. R²countryecoregionsbiomeavg. acc.
        RevGeoRQ.929 ± .001.896 ± .004.831 ± .004.685 ± .007.835.949 ± .000.723 ± .011.852 ± .012.841
        HierID baseline.515 ± .036.764 ± .004.722 ± .005.675 ± .002.669.851 ± .008.465 ± .002.711 ± .007.676
    • Additionally, embeddings derived from an LLM like Qwen3 capture semantics about the place itself. We have included some now experiments (provided in newly uploaded supplementary material) that demonstrate this via the following examples:
      • Language names align with the cities where the language is widely spoken.
      • City names cluster by country by default, and can cluster by other attributes (like size of the city) when prompted.
      • Place names align with descriptions of their physical setting.
      • Pls see newly submitted supplmenetary file for detailed explanations and figures about the above.
  • W2: "may bias the learned representations toward boundary-aware geographic characteristics while weakening their ability to capture other important spatial properties, such as environmental similarities." and "The authors should provide a more in-depth analysis of these trade-offs rather than relying primarily on aggregate performance improvements."
    • We agree that a detailed analysis of the trade-off between naturally occurring characteristics vs. admin boundary induced characteristics can provide deeper insights. We thank the reviewer for this suggestion, and will plan to include this as future work. We would like to highlight the following existing parts of our paper that partially address this:
      • Line 252: “ Results are mixed on smoother targets such as air temperature and elevation”
      • Line 468-470: "A limitation…. distinct locations within the same administrative boundaries may map to identical strings, even when they are far apart or differ substantially in local characteristics"
      • Line 470-472: "The hierarchical discretization of our model can also obscure non-local similarities, particularly between geographically distant regions on different continents."

(contd. below)

AuthorsOpenReview ↗

Rebuttal response to Reviewer meqg (2/n)

  • W3: "relies exclusively on simple administrative attributes without exploring richer semantic information, such as climatic or socioeconomic characteristics. This potentially limits the diversity and expressiveness of the learned representations."
    • Encoding place names using LLMs helps us capture richer characteristics that go beyond simple place names and admin boundaries. Examples:
      • The learned semantic IDs capture cultural-regional distinctions like South Asia, the Arabian Gulf, and so on (Line 405 Sec. 6.1). None of these terms are included in explicit place names data. These are emergent region-cultural subgroups that are learned emergent properties.
      • To show the kinds of rich semantic information contained in LLM embeddings of place names (which are our reconstruction target), we have included some now experiments (provided in newly uploaded supplementary material) with the following examples:
        • Language names align with the cities where the language is widely spoken.
          • We embedded six diverse language names (Hindi, English, Arabic, Spanish, Swahili, Mandarin). For each language, we picked ten well-known cities where that language is the most widely spoken. The 10-nearest neighbors for each language embedding (using cosine similarity) are almost always the cities where this language is the most widely spoken. Precision@10 for language to city retrieval was found to be 0.97.
          • An additional experiment was performed where “distractor” keywords were added by introducing 6 additional language embeddings and 12 city embeddings with two cities per distractor language. Additionally, 8 city embeddings were introduced that don’t match to any language in the main 6 cities or the distractor 6 cities. Language-to-city P@10 drops only marginally from 0.97 to 0.95. (chance: 0.17 without distractors, 0.13 with them).
        • City names cluster by country by default, and can cluster by other attributes when prompted.
          • 5 countries and 9 cities per country were chosen. Of the 9 cities, 3 are large cities, 3 medium-sized, and 3 small towns.
          • Default clustering by country: Performing t-SNE on these shows that the cities form five tight groups, one for each country. For every city, its 5 nearest neighbors (by cosine similarity) are all from the same country.
          • Prompted clustering by city size: Qwen3-Embedding accepts an optional instruction that tells the model what the embedding will be used for. The instruction is put in front of the text as Instruct: <instruction>\nQuery: <text>. We re-embedded the same 45 place names with the instruction "Given a place name, classify whether it is a big city, a mid-sized city, or a small town." With this instruction, the t-SNE plot reorganizes by size instead of country. All 15 big cities, from all five countries, merge into one cluster. The mid-sized cities and small towns share a second region, with a relatively coarser sub-grouping of medium-sized cities in one cluster, and another cluster of small towns.
        • Place names align with descriptions of their physical setting.
          • We embedded four short descriptions of physical settings: "warm coast", "cold mountain", "hot desert" and "tropical rainforest". For each description, we picked 15 well-known cities that fit it, 60 cities in total. Each city was then assigned to the description it is most similar to. This gave the correct description for 56 of the 60 cities (93%; chance is 25%).
          • On a world map, the assignments fall where one would expect. Desert cities cover the Sahara, Arabia, the US Southwest and central Australia. Rainforest cities cover the Amazon, the Congo basin and Borneo. Mountain cities cover the Alps, the Rockies, the Andes and the Himalayas. Coastal cities line the Mediterranean, Florida, California and Queensland.
        • Complete details and figures explaining the above can be found in the newly submitted supplementary PDF file.

(contd. below)

AuthorsOpenReview ↗

Rebuttal response to Reviewer meqg (3/n)

  • W4: "...the paper lacks ablation studies examining the contributions of different administrative attributes (e.g., city, region, country, and continent). It remains unclear whether coarse-grained information, such as country labels alone, could achieve comparable performance."
    • We thank the reviewer for this suggestion. While we did not run a full per-attribute ablation, the levels of the hierarchy play complementary roles, and removing any of them weakens the supervision in a predictable way:
      • Coarse levels alone cannot separate nearby locations. Dropping finer levels (city, region) merges all locations within a unit into a single target, removing the signal for within-unit variation in climate, elevation or population density that our downstream tasks measure. For example, our full reverse-geocoded strings yield approx. 21K unique targets on S2-100K, versus approx. 250 with country names alone.
      • Fine levels alone lack context. Dropping coarser levels (region, country) leaves names that are ambiguous or poorly represented. Well-known places (e.g., "New York") are well encoded by the Qwen3 embeddings, but less-known places improve noticeably when their admin1 and country names are appended. The hierarchy also disambiguates shared names: "Paris" alone embeds close to Paris, France, whereas "Paris, Texas, United States of America" correctly represents the Texas city.
    • The full hierarchy therefore combines the specificity of fine-grained names with the context of coarse ones. We agree that a systematic per-attribute ablation would strengthen this argument and will include it as future work.
  • Q1: "What motivates the choice of RQ-VAE and Qwen3-Embedding? Have the authors considered other alternatives, such as learnable categorical embeddings or direct embedding alignment?"
    • We thank the reviewer. We considered both alternatives:
      • Learnable categorical embeddings. As shown in our response to [W1] above, RQ-VAE + Qwen3 outperforms categorical embeddings trained on the same reverse-geocoded labels. Categorical IDs treat each administrative unit as an independent symbol and carry no information beyond the label itself.
      • Why RQ-VAE. Its residual codebooks yield coarse-to-fine Semantic IDs (Sec. 6.1), including groupings such as "South Asia" that never appear in the reverse-geocoded strings.
      • Why Qwen3-Embedding. New experiments (submitted via new a new supplementary file, and explained above in [W3]) shows that its place-name embeddings encode knowledge beyond the label itself: language names retrieve the cities where they are most widely spoken (P@10 = 0.97, or 0.95 with distractors), city names cluster by country by default and by city size when instructed, and cities match descriptions of their physical setting (93% vs. 25% chance).
  • Q2-a: "How accurately do the learned embeddings capture actual administrative boundaries?"
    • Please see Sec. 6.1 (lines 411–423), where we measure how well the learned Semantic IDs predict administrative attributes at unseen locations. Accuracy rises monotonically with prefix depth for every attribute (Fig. 2, right): continent is largely fixed by the first code (83%) and country by the first two (70%), while admin1 and city keep improving up to the full Semantic ID (61% and 33%). The lower admin1 and city accuracies reflect resolution rather than missing information (explored further in Appedix E).
    • Please also see Table 1 and Sec. 6.2: the continuous embeddings, which feed the downstream tasks, capture boundaries more precisely than this discrete view. On country classification, an MLP probe reaches 94.6% accuracy with SatCLIP-TextRQ and 94.9% with RevGeoRQ, versus 90.7% for native SatCLIP (Table 1). Qualitatively, the similarity fields in Fig. 3 (Sec. 6.2) align with country borders and taper across the subcontinent and continent, whereas native SatCLIP fragments within countries.

(contd. below)

AuthorsOpenReview ↗

Rebuttal response to Reviewer meqg (4/n)

  • Q2-b: "Which types of downstream tasks or geographic scenarios benefit most from boundary-aware supervision, and in which cases might it introduce undesirable biases or degrade performance?"
    • Please see Sec. 4.2.1, Sec. 6.2, Sec. 7 and App. F.
      • Benefits: Tasks aligned with administrative units benefit most. TextRQ improves population prediction for every backbone and country classification for three of four (App. F), and improves top-1 iNaturalist geo-prior accuracy for all backbones (Table 2).
      • Degradation: Tasks that vary within an administrative unit can degrade, such as California Housing and, for some backbones, air temperature and elevation (App. F).
      • Biases: Coverage is coarser in sparsely populated regions, and boundary-aware embeddings miss long-range physical similarities, such as between the Himalayas and the Andes, that image-trained encoders capture (Sec. 6.2, Sec. 7).
      • Mitigation: Concatenating with native embeddings recovers much of this, outperforming Native in 24 of 28 comparisons (Sec. 4.2.1).
  • Q3: "Why does the proposed framework rely exclusively on a relatively simple reverse-geocoding strategy? Have the authors considered extracting richer geographic semantics to improve representation diversity and generalizability?"
    • We thank the reviewer for this suggestion. We chose reverse geocoding because it is a simple, cheap and scalable signal, needed only offline during pretraining; at inference, the encoder takes raw coordinates directly. Enriching the strings with additional context as suggested, e.g., the nearby places used in GeoLLM (Manvi et al.), would add substantial cost at our pretraining scale. Moreover, richer semantics are already partly captured implicitly: the pretrained LLM encoder brings in knowledge beyond the place names themselves, such as associated languages, city size and physical setting (see our response to W3 and App. G). Since our framework accepts any text target, richer descriptions can be swapped in without changing the method, which we see as promising future work.
  • Q4: "How does the performance change when different levels of administrative information are used for supervision"
    • Addressed via response to W4 above.
Reviewer_eaf9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to learn location encoders that are aware of administrative boundaries. The dominant paradigm in location encoding (e.g. AEF, RANGE, etc.) is to reconstruct some kinds of primarily visual information, which are intuitively relatively uncorrelated with underlying semantic regions in the world. The paper argues that it is useful to learn human-percieved boundaries, and accomplishes this by aligning to text produced via reverse geocoding. Results show that learning administrative boundaries at multiple graularities is useful both by itself for targets relating to borders and is also useful in addition to standard location embedings.

Critical strengths

The core intuition is nice. Things like population vary with administrative boundaries while things like climate don't. It's also not always possible to perceive boundaries from imagery, which would make the paper's contribution generally complementary.

Training over different encoders is nice and shows transferability.

Critical weaknesses

Relevant work: G3: an effective and adaptive framework for worldwide geolocalization using large multi-modality models UNIGEOCLIP: Unified Geospatial Contrastive Learning Sat2Cap: Mapping Fine-Grained Textual Descriptions from Satellite Images

Are all relevant papers that align a text representation space to imagery. As such, the resulting textual representations are relevant baselines to compare the performance of this method.

The role of the information contained in the textual embeddings is not sufficiently discussed or ablated. The paper would benefit from more discussion of what aligning representations of place names does (e.g. what information is in place name embeddings) and qualify this with ablations on different LLMs for embedding creation and observe the results.

Should be noted that the label construction procedure seems to use a nearest lookup that would essentially Voronoi partition the world which is not really assured to match admin boundaries. Its unclear if this leaves performance on the table, and also seems like could be fixed relatively simply (the PIGEON geolocalization paper solves this problem with GADM)

More minor:

  • The figures are not visually appealing, difficult to read with text either being obstructed or too small.
  • Climplicit is labeled as ReSIREN in more than one place (eg table 5)
  • GeoLLM citation is messed up
  • More discussion about the multi-seed procedure would be nice
Questions for authors

Please address weaknesses.

In addition: Why is the last token state used from Qwen3-Embedding? The model produces a single embedding as an output. How do gradients flow through the discrete code lookup?

4 comments
AuthorsOpenReview ↗

Rebuttal response to Reviewer eaf9 (1/n)

We would like to thank Reviewer eaf9 for their detailed and thoughtful review. Please find our responses and improvements below:

  • Missing relevant work - thank you for pointing this out, we will add these to the applicable sections

  • “Are all relevant papers that align a text representation space to imagery. As such, the resulting textual representations are relevant baselines to compare the performance of this method.”

    • Work already considered:

      • LLMGeovec and GeoLLM (included in our literature review) - these require map text at inference time. To keep our work focused, we centered specifically on “pure location encoders” i.e. those that only need a lat-long pair at inference time
    • Newer work considered during rebutall phase:

      • CaLLiPer / CaLLiPer+ uses text from POI categories, encodes it using an LLM, and then performs alignment with corresponding geographic coordinates through contrastive learning; trained only on London data, so it does not transfer well to global tasks. Sharing results below for CaLLiPer vs. our model; CaLLiPer+ model weights have not been released.
      Modelair tempelevationpopulationcali housingavg. R²countryecoregionsbiomeavg. acc.
      RevGeoRQ.929 ± .001.896 ± .004.831 ± .004.685 ± .007.835.949 ± .000.723 ± .011.852 ± .012.841
      CaLLiPer−.780 ± .015−2.452 ± .423−1.459 ± .384−.553 ± .029−1.311.047 ± .006.010 ± .001.125 ± .002.061

      The above result may not be included in our final manuscript due to poor global transferability of this model.

      • LE4Mob - this uses textual descriptions of nearby POIs, encodes them using a pretrained LLM, and aligns coordinates with the text embeddings through contrastive learning while incorporating an additional loss to preserve geographic distances. No weights released. Will add to related works.
      • AddressBind - aligns coordinates to street addresses that are encoded via a text encoder. No publicly released weights or repo. Will add to related works.
      • UniGeoCLIP - this uses text descriptions, aerial imagery, street-view imagery, and digital surface models to supervise a location encoder, aligning all five modalities (including geographic coordinates) into a shared embedding space through contrastive learning. No weights released (as of Nov. 10 2026). Will add to related works.

(contd. below)

AuthorsOpenReview ↗

Rebuttal response to Reviewer eaf9 (2/n)

  • “The paper would benefit from more discussion of what aligning representations of place names does (e.g. what information is in place name embeddings)”
    • We’ve done some additional work around analyzing the information contained in place name embeddings. This is provided via an additional file uploaded now as supplementary material. We plan to add this to the final manuscript as well.
    • A summary of what we found (supporting figures in the supplementary file mentioned):
      • Language names align with the cities where the language is widely spoken.
        • We embedded six diverse language names (Hindi, English, Arabic, Spanish, Swahili, Mandarin). For each language, we picked ten well-known cities where that language is the most widely spoken. The 10-nearest neighbors for each language embedding (using cosine similarity) are almost always the cities where this language is the most widely spoken. Precision@10 for language to city retrieval was found to be 0.97.
        • An additional experiment was performed where “distractor” keywords were added by introducing 6 additional language embeddings and 12 city embeddings with two cities per distractor language. Additionally, 8 city embeddings were introduced that don’t match to any language in the main 6 cities or the distractor 6 cities. Language-to-city P@10 drops only marginally from 0.97 to 0.95. (chance: 0.17 without distractors, 0.13 with them).
      • City names cluster by country by default, and can cluster by other attributes when prompted.
        • 5 countries and 9 cities per country were chosen. Of the 9 cities, 3 are large cities, 3 medium-sized, and 3 small towns.
        • Default clustering by country: Performing t-SNE on these shows that the cities form five tight groups, one for each country. For every city, its 5 nearest neighbors (by cosine similarity) are all from the same country.
        • Prompted clustering by city size: Qwen3-Embedding accepts an optional instruction that tells the model what the embedding will be used for. The instruction is put in front of the text as Instruct: <instruction>\nQuery: <text>. We re-embedded the same 45 place names with the instruction "Given a place name, classify whether it is a big city, a mid-sized city, or a small town." With this instruction, the t-SNE plot reorganizes by size instead of country. All 15 big cities, from all five countries, merge into one cluster. The mid-sized cities and small towns share a second region, with a relatively coarser sub-grouping of medium-sized cities in one cluster, and another cluster of small towns.
      • Place names align with descriptions of their physical setting.
        • We embedded four short descriptions of physical settings: "warm coast", "cold mountain", "hot desert" and "tropical rainforest". For each description, we picked 15 well-known cities that fit it, 60 cities in total. Each city was then assigned to the description it is most similar to, using calibrated similarity (see the note on calibration). This gave the correct description for 56 of the 60 cities (93%; chance is 25%).
        • On a world map, the assignments fall where one would expect. Desert cities cover the Sahara, Arabia, the US Southwest and central Australia. Rainforest cities cover the Amazon, the Congo basin and Borneo. Mountain cities cover the Alps, the Rockies, the Andes and the Himalayas. Coastal cities line the Mediterranean, Florida, California and Queensland.
      • Note on calibration. The embedding space produced by Qwen3, like many other LLMs, is strongly anisotropic: almost any two strings have a cosine similarity of about 0.5, so a good match often scores only slightly higher than a poor one. To correct for this, when we assign each place to a label, we first subtract that label's average similarity over all places in the experiment.. Rankings within a single label, such as the language-to-city top-10 retrieval and the nearest-neighbor counts, are not changed by this step, so those results use plain cosine similarity.
  • “ablations on different LLMs for embedding creation”
    • This is already partially addressed in Line 350-351 in Table 3, and line 358: “Replacing Qwen-0.6B with Qwen-4B provides further but modest gains.”
    • Additionally, we had tried a few LLMs that are comparable to the Qwen3- Embedding models that are our primary text encoders, like the Nemotron-3-Embed-1B. We did not find substantial perf differences for these comparable models.
    • Some additional new experimentation reveals that comparing the Qwen-0.6B performance (line 347) with smaller models shows degraded performances with models that have fewer parameters / are trained on limited data
      • Qwen-0.6B scores .729/.707 R^2 / acc.
      • BabyLM GPT-2 (100M words): .695/.671
      • GPT-1 (fiction books): .675/.670.
      • BabyLM GPT-2 (100M words): .695/.671.
      • TinyStories-33M: .564/.597. contd. below
AuthorsOpenReview ↗

Rebuttal response to Reviewer eaf9 (3/n)

  • "label construction procedure seems to use a nearest lookup that would essentially Voronoi partition the world which is not really assured to match admin boundaries”
    • We thank the reviewer for this observation. The reverse_geocoder library assigns each coordinate the labels of its nearest GeoNames populated place, which implicitly induces a Voronoi partition. Near administrative boundaries, a point can therefore inherit the labels of a neighboring unit. This error is concentrated along boundaries, and the affected band widens where populated places are sparse.
    • Our method is agnostic to the reverse geocoder, since it only consumes the returned text, so a GADM point-in-polygon lookup (as in PIGEON) is a drop-in replacement. However, this would reduce rather than eliminate misalignment: GADM polygons have country-dependent digitization error, and coastal or gap points still need a nearest-polygon fallback. We will discuss this implicit tessellation as a limitation and direction for future work.
    • The larger mismatch we observed was with continuous natural features such as elevation and temperature, which do not follow administrative boundaries at all. We investigated this through the following:
      • Appendix F quantitatively examines this via loss of spatial granularity
      • Line 252: “ Results are mixed on smoother targets such as air temperature and elevation”
      • Line 468-471: A limitation…. distinct locations within the same administrative boundaries may map to identical strings, even when they are far apart or differ substantially in local characteristics;
      • Line 472-474: “performance depends on reverse-geocoding quality. More precise geocoders may recover finer distinctions, but coverage remains inherently limited in large, sparsely populated regions where neighborhood- and locality-level boundaries are coarse or absent.
  • "Why is the last token state used from Qwen3-Embedding? The model produces a single embedding as an output."
    • We thank the reviewer and apologize for the unclear wording. The last-token state is the model's single output embedding. As described in the Qwen3 Embedding paper (Sec. 2, "Embedding Models"; https://arxiv.org/pdf/2506.05176): "The final embedding is derived from the hidden state of the last layer corresponding to this [EOS] token." Because of causal self-attention, this final token attends to all preceding tokens and therefore summarizes the full input string.
    • The only difference from the default usage is that we take this embedding before the final L2 normalization ("unnormalized" in Sec. 3.2.2), because normalized targets led to degenerate solutions in our experiments. This is like due to the fact that the RQVAE reconstruction loss is the squared L2 distance between the target and reconstructed text embeddings, L_recon = ||t_i − t̂_i||² (Sec. 3.2.3). We will revise the text to make this clearer.
  • "How do gradients flow through the discrete code lookup?"
    • The stop-gradient operator is used for these layers/parts of the models. Addressed in line 189-190: Here sg[·] denotes the stop-gradient operation (Van Den Oord et al., 2017).
  • Minor corrections section: we thank the reviewer for their detailed review! We will fix all of these in the final manuscript

Thank you for your detailed response during a busy rebuttal period.

Generally, it's still unclear what information is expected to be useful from the textual embeddings. It's also unclear what benefits the method provides over storing the polygons associated with admin boundaries along with the teacher text embedding from Qwen. This would be less params and is directly the teacher signal the model is aligned to. The author(s) of the paper should think about the motivation of learning from admin boundary representations generally to revise the formulation. I absolutely think that it's worth thinking about both human-derived semantic insights that come from text along with the general paradigm of learning from large-scale imagery. These are likely complementary, and as people we understand the world through such semantic insights about places without observing significant amounts of aerial imagery, and have learned about geographic phenomena through textbooks without recovering them from implicit neural representations. However, the paper at present doesn't convince me that this is the right direction for this. I encourage thinking about extensions like enriching the text by finding geolocated text from sources like wikipedia, or as an augmentation to more spatially grounded architectures like Tessellating the Earth [ECCV'26].

I will keep my score. Again, thank you for your effort during the rebuttal.

Reviewer_UcKTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a method to explicitly train location encoding with administrative supervision. In short, they add an auxiliary task to align the learned embedding with frozen text embedding of the hierarchical place addresses using an RQ-VAE. The authors reported minor improvements on downstream tasks.

Critical strengths

S1. The authors propose to explicitly train location encoding with administrative supervision.

Critical weaknesses

W1. The critical missing part is a tree-structure supervision as a basic alternative to what the paper proposes, i.e. the text-based reverse-geocoded string. I understand that the authors are motivated that the hierarchical address of the looked-up location (in GIS, such geo-coding is usually a database with human-calibrated administrative levels) will serve as a quasi-distance signal, e.g. if two locations are in neighboring counties, their (admin2, country, continent) tuple will look all the same. But the key question is: what is unique by using texts instead of the IDs of administrative levels that already exist in the geocoding database? Such a database is by itself a well organized tree structure, where a tree-distance can be deterministically calculated as a training signal without using any models. This is the most natural administrative information one should use in training a location encoder.

W2. To move one step forward: the authors should do the ablation study by replacing place names with number IDs and do the same experiments they did in the paper. That will isolate the effect of using HIERARCHY and the effect of exposing PLACE NAMES (e.g. Beijing sounds linguistically distant from New York, even without a complete hierarchy).

Minor concerns: the reported improvement is very marginal (~0.5%) in many cases.

Questions for authors

Q1. Have you ever tried to use the tree-distance on the administrative hierarchy of places in place of your frozen text embedding as the reconstruction target?

Q2. Is there any difference between using the reverse-geocoded string with explicit place names and using a string of place IDs? That will isolate whether the improvement comes from the hierarchy or simply from knowing more about the place names.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Rebuttal response to Reviewer UcKT (contd. 3/n)

  • To show the advantages of using LLM based embedding as the reconstruction target, we’ve done some additional work around analyzing the information contained in place name embeddings. This is provided via an additional file uploaded now as supplementary material. We plan to add this to the final manuscript as well.
  • A summary of what we found (supporting figures in the supplementary file mentioned):
    • Language names align with the cities where the language is widely spoken.
      • We embedded six diverse language names (Hindi, English, Arabic, Spanish, Swahili, Mandarin). For each language, we picked ten well-known cities where that language is the most widely spoken. The 10-nearest neighbors for each language embedding (using cosine similarity) are almost always the cities where this language is the most widely spoken. Precision@10 for language to city retrieval was found to be 0.97.
      • An additional experiment was performed where “distractor” keywords were added by introducing 6 additional language embeddings and 12 city embeddings with two cities per distractor language. Additionally, 8 city embeddings were introduced that don’t match to any language in the main 6 cities or the distractor 6 cities. Language-to-city P@10 drops only marginally from 0.97 to 0.95. (chance: 0.17 without distractors, 0.13 with them).
    • City names cluster by country by default, and can cluster by other attributes when prompted.
      • 5 countries and 9 cities per country were chosen. Of the 9 cities, 3 are large cities, 3 medium-sized, and 3 small towns.
      • Default clustering by country: Performing t-SNE on these shows that the cities form five tight groups, one for each country. For every city, its 5 nearest neighbors (by cosine similarity) are all from the same country.
      • Prompted clustering by city size: Qwen3-Embedding accepts an optional instruction that tells the model what the embedding will be used for. The instruction is put in front of the text as Instruct: <instruction>\nQuery: <text>. We re-embedded the same 45 place names with the instruction "Given a place name, classify whether it is a big city, a mid-sized city, or a small town." With this instruction, the t-SNE plot reorganizes by size instead of country. All 15 big cities, from all five countries, merge into one cluster. The mid-sized cities and small towns share a second region, with a relatively coarser sub-grouping of medium-sized cities in one cluster, and another cluster of small towns.
    • Place names align with descriptions of their physical setting.
      • We embedded four short descriptions of physical settings: "warm coast", "cold mountain", "hot desert" and "tropical rainforest". For each description, we picked 15 well-known cities that fit it, 60 cities in total. Each city was then assigned to the description it is most similar to, using calibrated similarity (see the note on calibration). This gave the correct description for 56 of the 60 cities (93%; chance is 25%).
      • On a world map, the assignments fall where one would expect. Desert cities cover the Sahara, Arabia, the US Southwest and central Australia. Rainforest cities cover the Amazon, the Congo basin and Borneo. Mountain cities cover the Alps, the Rockies, the Andes and the Himalayas. Coastal cities line the Mediterranean, Florida, California and Queensland.
    • Note on calibration. The embedding space produced by Qwen3, like many other LLMs, is strongly anisotropic: almost any two place-name strings have a cosine similarity of about 0.5, so a good match often scores only slightly higher than a poor one. To correct for this, when we assign each place to a label, we first subtract that label's average similarity over all places in the experiment.. Rankings within a single label, such as the language-to-city top-10 retrieval and the nearest-neighbor counts, are not changed by this step, so those results use plain cosine similarity.

Thanks for the updated results. I'd keep my rating.

Reviewer UcKT, did we sufficiently addressed your concern? May I ask why you’re choosing to keep your rating?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.