TLDR: A new benchmark, “Epidemiology of Large Language Models,” evaluates if LLMs can accurately reflect real-world probabilistic distributions, specifically observational knowledge (Layer 1 of Pearl’s Causal Hierarchy). The study found that current LLMs perform poorly in understanding these statistics across various domains, scoring an average of 22/100 in low-dimensional and 17/100 in high-dimensional tasks. This suggests limitations in their ability to reason about interventions or counterfactuals, with fine-tuning and web-based retrieval showing limited benefits.
Artificial intelligence systems, particularly Large Language Models (LLMs), have shown incredible advancements, but a new study delves into a crucial question: how well do these models truly understand the probabilistic nature of the real world? While much attention has been paid to their ability to recall facts, researchers are now exploring whether LLMs can grasp the underlying statistical distributions that describe populations and phenomena.
The paper, titled “EPIDEMIOLOGY OF LARGE LANGUAGE MODELS: A BENCHMARK FOR OBSERVATIONAL DISTRIBUTION KNOWLEDGE,” introduces a novel benchmark to specifically test this capability. Authored by Drago Pleˇcko, Patrik Okanovi´c, Torsten Hoefler, and Elias Bareinboim, the research highlights a critical distinction between factual knowledge (e.g., “what is the capital of England?”) and probabilistic knowledge (e.g., “what is the sex of a computer science graduate in the US?”).
The Challenge of Probabilistic Knowledge
LLMs are trained on vast amounts of text, leading some to believe they might naturally internalize real-world probability distributions. However, classical statistical theory, particularly the “curse of dimensionality,” suggests that learning distributions in high-dimensional spaces is inherently difficult. This tension between optimism about LLMs as universal approximators and statistical caution forms the core motivation for this research.
The study focuses on “observational probabilistic knowledge,” which corresponds to Layer 1 of Pearl’s Causal Hierarchy (PCH). The PCH outlines different levels of cognitive capabilities: Layer 1 for observation (seeing correlations), Layer 2 for intervention (understanding effects of actions), and Layer 3 for counterfactual reasoning (imagining alternative realities). The Causal Hierarchy Theorem states that if a model lacks observational knowledge (Layer 1), its ability to perform reliable inferences at higher layers (interventional and counterfactual) is also fundamentally limited.
Introducing the Epidemiology Benchmark
To evaluate LLMs, the researchers developed a benchmark that can be thought of as an “epidemiology” for AI – the knowledge of what falls upon populations. This benchmark utilizes 10 large, publicly available datasets from the United States, covering diverse domains such as economics (American Community Survey, Survey of Consumer Finances), health (National Health and Nutrition Examination Survey, Behavioral Risk Factor Surveillance System, Medical Expenditure Panel Survey, National Survey on Drug Use and Health), education (Department of Education), labor (Department of Labor BLS Data), social behavior (General Social Survey), and crime statistics (FBI Arrest Statistics).
The benchmark creates tasks that probe LLMs about conditional distributions. For example, a task might ask, “What is the sex of a person who graduated with a degree in physical sciences?” The model’s predicted probabilities for answers (e.g., male/female) are then compared against the true distributions derived from the datasets. The evaluation considers both low-dimensional tasks (conditioning on one variable) and high-dimensional tasks (conditioning on multiple variables).
Key Findings: LLMs Fall Short
The results were quite revealing: current LLMs generally perform poorly in understanding real-world observational distributions. The best models scored an average of only 22 out of 100 points in low-dimensional settings and a mere 17 out of 100 in high-dimensional settings. While they performed better than random guessing, their scores indicate a significant shortfall from accurately reflecting real-world statistics.
Even advanced closed-source models like OpenAI’s GPT-4.1 and o4-mini, while outperforming open-weight models, still showed limited capabilities. Interestingly, for some tasks, even Retrieval-Augmented Generation (RAG), which allows models to search the web for information, did not lead to improved performance, suggesting that simply having access to data isn’t enough for internalizing these complex distributions.
The study also explored whether fine-tuning LLMs on synthetic data generated from the ground truth distributions could help. The findings indicated only limited improvements, particularly for low-dimensional tasks using likelihood-prompting, but not consistently across all settings or models.
Also Read:
- New Benchmark Reveals LLMs Struggle to Grasp Deep Human Values, Favoring Surface Preferences
- Understanding LLM Reliability: A Hierarchical Imprecise Probability Framework
Implications for AI Development
These findings have significant implications. If LLMs struggle to internalize basic observational distributions, their ability to perform more complex causal inferences – such as understanding the effects of interventions or reasoning about counterfactual scenarios – is likely to be severely hampered. This suggests that current approaches to using AI models for causal inference may need more sophisticated methods.
The research underscores the need for continued work on how to imbue AI systems with a deeper, more robust understanding of the probabilistic world. The benchmark itself is designed to be reproducible and extensible, encouraging future research to evaluate more models and datasets. For more details, you can read the full research paper here.
