Locations-to-Location Translation (LTL)

• Locations-to-Location Translation (LTL) is a class of problems that maps compositional natural language descriptions and mobility sequences to formal geospatial outputs.
• It utilizes methods such as geoparser augmentation, chain-of-thought reasoning, and hierarchical tokenization to derive precise bounding boxes and semantic location tokens.
• Recent advances demonstrate significant improvements in geocoding accuracy and mobility prediction metrics, validating LLM-based techniques over traditional baselines.

Locations-to-Location Translation (LTL) refers to a class of problems in which systems map from complex, contextually described or sequential collections of locations—ranging from compositional references in natural language to trajectories in urban mobility data—into formal geospatial outputs such as bounding boxes or discrete location embeddings. Recent advances leverage LLMs to enhance both geocoding performance for text-based spatial descriptions and semantic modeling for mobility sequences, yielding substantive improvements over traditional baselines (Masis et al., 9 Oct 2025, Chen et al., 8 Jun 2025).

1. Formal Problem Definition

LTL encompasses settings where the input is not a canonical location identifier but a composition of locations, either in natural language (e.g., “midway between A and B”), or as a temporal sequence in mobility analytics. The geocoding variant, as formally defined in (Masis et al., 9 Oct 2025), receives as input a compositional natural-language description $d$ denoting an unnamed location $\ell$, and a set $M = \{ m_1, ..., m_k \}$ of referenced toponyms. The system’s objective is to predict a rectangular bounding box grounding $$b_\ell = (\text{lon}_\text{min}, \text{lat}_\text{min}, \text{lon}_\text{max}, \text{lat}_\text{max})$$.

The computation is decomposed into:

• Recaller $g$: $M \rightarrow \{q_i\}$, retrieving for each $m_i$ a geospatial representation (e.g., center coordinates).
• Reasoner $f$: $(d, \{q_i\}) \rightarrow b_\ell$, integrating the description and retrieved location context to yield the bounding box.

In the trajectory-based instantiation (e.g., (Chen et al., 8 Jun 2025)), the system predicts the next or missing location in a sequence, using learned semantic location tokens mapped from heterogeneous descriptors.

2. LLM-based Approaches in LTL

A. LLMs for Grounding Contextual Location References

The “Coordinates from Context” approach investigates prompt-based and fine-tuned LLM strategies for LTL geocoding, distinguishing between:

• Direct: LLM infers $b_\ell$ solely from $\ell$0, without explicit context.
• Geoparser-Augmented: External geocoding of each $\ell$1 yields $\ell$2 as input to the LLM, enabling explicit spatial reasoning.
• Chain-of-Thought Decomposition: Prompts scaffold intermediate reasoning, requesting explicit listing of min/max coordinates from the context.

Few-shot learning with two-shot templates is employed for both geospatial knowledge (predict point from name) and geospatial reasoning (output bounding box from description and context), with model-centric variations in context utilization and fine-tuning (LoRA, small or large LLMs).

B. Semantic Location Tokenization and Alignment for Mobility Analytics

QT-Mob introduces a hierarchical discrete tokenization of locations via Residual-Quantization VAE (RQ-VAE). Given a location descriptor $\ell$3, a pre-trained LLM encodes $\ell$4 into $\ell$5, which is then quantized into $\ell$6 discrete codes using hierarchical codebooks, yielding compact, semantic tokens $\ell$7 and quantized embedding $\ell$8. These are integrated as vocabulary additions for the LLM, supporting token-level representations of arbitrary locations.

Multi-objective fine-tuning aligns tokenized locations with the LLM's internal semantics, using objectives such as next-location prediction, masked trajectory recovery, and bidirectional text-token alignment.

3. Key Architectures and Algorithms

Approach	Input	Output / Target
Geocoding LTL (Masis et al., 9 Oct 2025)	$\ell$9, $M = \{ m_1, ..., m_k \}$0	Bounding box $M = \{ m_1, ..., m_k \}$1
QT-Mob LTL (Chen et al., 8 Jun 2025)	$M = \{ m_1, ..., m_k \}$2	Next location code sequence $M = \{ m_1, ..., m_k \}$3

Core Formulas

• Bounding Box Evaluation (Masis et al., 9 Oct 2025):
  • Haversine centroid distance:

  $M = \{ m_1, ..., m_k \}$4 - Area-overlap: $M = \{ m_1, ..., m_k \}$5 with respective numerator/denominator definitions.

• Semantic Tokenization (QT-Mob) (Chen et al., 8 Jun 2025):
  • Residual-Quantization VAE produces tokens:

  $M = \{ m_1, ..., m_k \}$6

  $M = \{ m_1, ..., m_k \}$7 - Next-location modeling with beam search and token-level constraints.

4. Evaluation Protocols and Benchmarks

A. Datasets

• GEOCODE (Masis et al., 9 Oct 2025): 1,000 compositional descriptions from Wikipedia, with gold-standard bounding boxes from OSM.
• Mobility Datasets (Chen et al., 8 Jun 2025): NYC, SG, and CE city datasets for trajectory-based prediction and recovery.

B. Metrics

• Geocoding LTL: Coverage (% valid predictions), centroid distance error (Haversine, km), convex box area overlap (F1).
• Mobility LTL: Hit@K, N@K for next location; masked reconstruction for mobility recovery.

C. Baselines

• Knowledge (pure toponym-to-point by LLM).
• Reasoning (description plus context).
• Direct and geoparser-augmented pipelines.
• Grammar-Based Spatial Parser (GBSP).
• End-to-end LLM without external geocoding.

5. Experimental Results and Error Analysis

A. Geocoding LTL (Masis et al., 9 Oct 2025)

Model	Coverage	Dist (km)	Area F1
Qwen 72B geo-aug (oracle)	98.5%	36.2	0.251
FT-Qwen 14B geo-aug (Google)	90.8%	240.9	0.266
GBSP (Laparra & Bethard)	52.8%	—	0.240

LLMs exhibit high error as geospatial knowledge sources (200–330 km), but when supplied with explicit toponym context, error drops below 100 km and F1 surpasses classical parser baselines. Geo-augmented LLMs tend to over-generalize, yielding overly large bounding boxes (high recall, lower precision), and fine-tuned small LLMs can match off-the-shelf large models. Direct LLM approaches are less accurate, with Area F1 as low as 0.054 for Qwen 14B.

B. Mobility LTL (Chen et al., 8 Jun 2025)

Dataset	Model	Hit@1	Hit@5	N@5
NYC	MobilityLLM	0.1840	0.3944	0.2949
NYC	QT-Mob	0.2550	0.4073	0.3366

QT-Mob outperforms baseline MobilityLLM by 38.5% in Hit@1 and over 14% in N@5, with similar uplifts for recovery metrics (e.g., 0.2284 Hit@1 for mobility recovery vs. 0.1586 for LLM-SFT). These gains validate the efficacy of semantically aligned, token-based location modeling with LLMs.

6. Practical Considerations and Recommendations

• Geoparser Access: Geoparser-augmented pipelines with external geocoding and a small, fine-tuned LLM (Qwen 14B or Llama 8B) yield the best LTL geocoding results.
• Resource Limitations: Absent geoparsers, large off-the-shelf LLMs (Llama 405B) are preferable for direct LTL.
• Pure LLM Knowledge: Relying solely on LLM geospatial knowledge (without context) leads to prohibitively high coordinate errors.

7. Significance and Future Directions

Locations-to-Location Translation, as operationalized in recent work (Masis et al., 9 Oct 2025, Chen et al., 8 Jun 2025), demonstrates that LLMs, when properly prompted or semantically aligned, can reliably bridge between compositional references and geospatial formalisms. Key results establish that the major advances are not in the implicit geospatial knowledge present in LLMs, but in their capacity for structured reasoning when provided grounded context or well-aligned discrete location representations. This division underlines the need for careful externalization, tokenization, and alignment of spatial signals in future LLM-based geospatial and mobility systems. Further research may investigate joint optimization of recaller and reasoner modules, richer tokenization schemes, and broader coverage beyond compositional or sequential urban settings.