Spoiler: No, not on their own. But LGND has developed a way to help them (hint: Strabo).

Spatial thinking is the cognitive ability to visualize and reason about the shapes, sizes, and locations of objects and their relationships in space. Humans have a unique ability to express infinite spatial reasoning. We can think and reason about galaxies or subatomic particles that we have never seen.

Where LLMs fall short

Large language models do not experience or think about space as humans do. Even models that can look at images lack the lived, embodied sense of place that humans build from moving through the world. Instead, language models approximate spatial thinking by translating their textual understanding of the world to geometric understanding through tokens, statistical relationships, a bit of math, and some guesswork. 

For LLMs, space is a math problem hidden within text. That gets them surprisingly far, but it breaks down quickly on real geospatial questions, where the answer depends on what is physically happening on the ground at a particular place and time.

Tools help, but only partly

LLMs lean on tools. MCP integrations load structured information about the world into a context window, such as:

• the Google Maps MCP answers routing questions
• a geocoder turns unstructured address strings into structured coordinates
• a weather service reports past, present, or future conditions at a given location

Each tool provides an isolated fact about the Earth. The model still has to stitch those facts together. It has no view of the broader context around them, what the surrounding landscape looks like, how it has changed, or what other places resemble it.

LGND’s Approach

Last week we introduced Strabo, our geosimilarity engine, and named agentic reasoning as one of the use cases it enables. Here’s how that works, from the ground up:

• Our inference pipelines convert massive imagery datasets into earth embeddings: numerical representations of what is happening on the Earth’s surface.
• Strabo, our geosimilarity engine, indexes those earth embeddings across spatial and temporal scales, so it can serve exactly the context a question needs.
• Through MCP, Strabo gives the LLM structured, multi-scale context about the physical world, rather than a handful of disconnected facts, so it can actually reason about places.

Some tasks require global context, some require local, some require yearly, some require daily. You shouldn’t need to switch implementations to handle the differences. And because Strabo builds space and time into how it indexes and ranks embeddings, the LLM gets context that already reflects where and when things are.

The system behaves similarly to how a geospatial tile server delivers context to end users. The coffee shop down the street doesn’t show up in Google Maps when you are looking at the country scale, only when you zoom in.

See it in action: Earth Agent

To show what this looks like in practice, we built Earth Agent, a demo application that pairs Strabo's embeddings with reasoning models. You can ask questions about the Earth in plain language and watch the model pull in the spatial and temporal context it needs to answer them.

Try the demo: https://demo.lgnd.ai/earth-agent/

Take the first sample prompt in the demo: “Where around Phoenix has new housing gone up since 2020?”

If you ask an LLM this question without Strabo, the reasoning path you will most commonly see is a check of permit applications submitted or issued as a proxy for new housing. But a permit isn’t a house. With Strabo, the agent checks the imagery itself: it finds candidate sites, compares before-and-after images, and reports where construction actually happened.

In the image below, green markers are sites where the imagery confirms new housing. Red markers came up in the LLM search, but the imagery showed no new building, so the agent flags them as unconfirmed rather than counting them. The chat panel on the left explains each finding.

A second example makes the gap even clearer: “Show me burn scar of recent European wildfires and how the areas have recovered since.”

Without Strabo, an LLM’s response was:

With Strabo, the agent located burn scars from three recent fires (Evia 2021, Losacio 2022, and Evros 2023), pulled before-and-after imagery for each, and tracked them through 2025.

So, can LLMs think spatially? No, not on their own. But give them the right context at the right scale, and they can start reasoning about the physical world in ways text and scattered facts alone can't support. That's what Strabo is built to do.

What's next

Earth Agent isn't the product. It's a window into what becomes possible when LLMs have access to Strabo. Next week, Strabo becomes available on the LGND platform, so you can start building applications like this yourself.

In the meantime, try Earth Agent at https://demo.lgnd.ai/earth-agent/ and check back on our blog for launch details.