Imagine asking a computer to estimate the cost of a two-bedroom Airbnb in Newark, New Jersey, without access to any real estate data. Sounds impossible, right? New research demonstrates how Large Language Models (LLMs) can tackle these “guesstimation” problems by building ad hoc probabilistic models on the fly. It turns out that while LLMs might struggle with direct calculations, they possess a wealth of commonsense knowledge that can be harnessed for surprisingly accurate estimations. The key innovation lies in extracting this knowledge, not as rigid rules, but as flexible constraints on a probability distribution. For example, an LLM might know that location and property type significantly influence price. Researchers prompt the LLM to identify relevant variables (like price, location, property type, number of beds) and then to specify constraints on how these variables interact. These constraints, expressed as conditional probabilities (e.g., "Given a downtown location, what’s the probability of the property being an apartment?"), are then used to build a probabilistic model. This model is optimized to satisfy all the constraints simultaneously, effectively combining fragmented bits of knowledge into a coherent whole. Experiments on real-world datasets (Airbnb listings, time-use surveys, and values surveys) revealed that this approach can generate estimates comparable to directly prompting the LLM for an answer. While not always superior to direct prompting, the approach holds significant potential for complex questions where direct answers are difficult to obtain. The real power of this method lies in its ability to integrate diverse sources of commonsense reasoning. Future work could refine this process by incorporating LLM confidence scores to weigh constraints and allowing for more complex constraint types, ultimately leading to even more accurate and robust guesstimation capabilities in LLMs.