Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study investigates the adherence of language models (LLMs) to the principle of statistical self-consistency, specifically examining how well their estimates align with the law of total probability when aggregating conditional distributions. By employing binary trees to create increasingly fine-grained subpopulations, the authors prompt LLMs with descriptions of these subpopulations and compare the resulting estimates to population-level marginals. The findings reveal significant violations of consistency properties, highlighting a phenomenon termed the macro fallacy, where fine-grained estimates are often more accurate than direct population-level estimates, indicating a gap in how models utilize their subpopulation knowledge.
LLMs often fail to propagate subpopulation knowledge into aggregate estimates, revealing a critical inconsistency in their probabilistic reasoning.
In-context learning is commonly interpreted as a form of conditional inference, in which the prompt specifies a context and the model's output is treated as an estimate of the corresponding conditional distribution. If this interpretation holds, then LLM estimates should satisfy basic probabilistic identities. In particular, the law of total probability asserts that prior-weighted conditional distributions aggregate into population-level marginals over any valid partition of the population. In this work, we investigate to what extent LLM estimates adhere to this self-consistency principle. We use binary trees as an evaluation scaffold to recursively partition a population into increasingly fine-grained subpopulations. We then prompt LLMs with verbalized subpopulation descriptions in context, aggregate the resulting estimates back into population-level estimates, and compare them across partitions of varying granularity. Applying this protocol across problem domains and state-of-the-art frontier models, we show widespread violations of basic consistency properties. An in-depth study of persona prompting reveals a pattern we call the macro fallacy: estimates reconstructed from more fine-grained subpopulation responses are often better aligned with human reference data than direct population-level estimates. This effect persists across variations in tree structure and estimation task, and can be partially recovered through implicit prompting. Together, these findings suggest that models possess relevant subpopulation knowledge but do not reliably propagate it into aggregate estimates. This gap establishes statistical self-consistency as an unsaturated, reference-free criterion for evaluating LLMs.