Researchers built a system that estimates household wealth by pairing satellite photos with text, and it outperforms image-only approaches by a meaningful margin.
The study uses demographic survey data from African neighborhoods to test five prediction pipelines: a vision model alone, an LLM given only location and year, an AI search agent pulling live web text, a joint image-text encoder, and an ensemble of all four. The best-performing setup reached an R-squared of 0.77 on out-of-sample data, compared to 0.63 for satellite imagery alone. Notably, the LLM's internal knowledge — what the researchers call "artificial neural memory" — proved surprisingly useful when predicting wealth in countries and time periods the model had never seen. The AI search agent, however, added only marginal and inconsistent gains, which the authors describe as limited evidence for the idea that live-retrieved data introduces genuinely novel signal.
The gap between 0.63 and 0.77 is large enough to matter for development organizations that use poverty maps to allocate resources. The finding that an LLM's baked-in world knowledge generalizes across borders is the more interesting result — it suggests language models have absorbed enough socioeconomic texture from pretraining that they can stand in for expensive ground surveys, at least partially. The search agent underperforming expectations is a useful data point against hype around agentic retrieval as a universal improvement.
The team also releases a dataset of roughly 60,000 survey clusters, each linked to satellite images and LLM-generated descriptions — the kind of resource that will let other researchers stress-test these findings before anyone stakes a poverty-reduction budget on them.