MoRA:

Location representations powered by human mobility.

MoRA encodes locations across the entirety of China into compact 128-dimensional representations capturing their socio-economic and functional DNA by anchoring satellite imagery, 100M+ POIs, and demographics to a billion-edge mobility graph.

ICLR 2026
Similarity View
Shows the cosine similarity between the representation of a chosen hexagonal grid cell and the remaining cells.
Cosine Similarity
Lower Higher
Static similarity map of East China

01 — The question

Two paradigms for representing locations

There have gradually emerged two main paradigms in geospatial representation learning: one grounds locations in their physical attributes, while the other represents them through patterns of human activity that unfold within them.

Earth Observation

What does this place look like?

Satellites, radar, LiDAR. These build a digital twin of the planet's surface — land cover, terrain, built environment. Powerful for monitoring, capturing the physical state of locations.

Physical state

Human-Centric

What do people do here?

Mobility, demographics, points of interest. A train station and an adjacent park share near-identical pixels, but radically different human dynamics.

Functional meaning

MoRA's claim is fundamentally human-centric. MoRA posits that a location's latent meaning lies less in its intrinsic attributes and more in its functional relationships with other places, as revealed by human movement.

02 — MoRA's core insight

Mobility as the syntax of space

MoRA borrows its theoretical logic from LLMs: meaning emerges from how discrete tokens co-occur in context. H3 grid cells become spatial tokens, and human mobility sequences provide the context that reveals each region's latent functional meaning.

Large Language Models (LLMs)
student reads map train market morning bright question city answer walks home
The student reads a map before taking the train home
[Semantic Meaning]
Spatial Intelligence Model
+

03 — Model architecture

Mobility as backbone for multimodal alignment

By using a human mobility graph as the central backbone for multimodal alignment, MoRA ensures POI text, satellite imagery, and demographic signals are interpreted through the lens of fundamental human dynamics.

04 — Results

General-purpose benchmark & results

We construct a comprehensive benchmark of 9 general-purpose prediction tasks across 4 spatial scales (point, grid, county, city) in social and economic domains, explicitly emphasizing human-centric evaluation.

12.9%
Avg. improvement
9
Benchmark tasks
0.78
Avg. R2 across tasks
Hover over each task to see its description and spatial scale
MoRA (ours)
Avg R2 = 0.78
AlphaEarth
Avg R2 = 0.68
SatCLIP
Avg R2 = 0.49

05 — Try MoRA

Coordinates in, intelligence out

The distilled MoRA model compresses an entire national-scale multimodal pipeline into a single function call. No satellite tiles. No POI databases. No mobility logs. Just coordinates (longitude/latitude).

Input
31.2304° N
121.4737° E

One coordinate pair is enough. MoRA inference does not require satellite tiles, POI databases, mobility logs, or census tables at runtime.

Usage
model = load_distilled_mora(model_path)

coord = [121.4737, 31.2304] # lon, lat

emb = model.encode(coord)

A single `encode(coord)` call returns a ready-to-use location fingerprint.

Output
128-D embedding

Encodes mobility, POI, imagery, and demographics into one compact vector representation.

Read paper→ Code→