Study on Statistical Self-Consistency in LLMs

AlayaLab · hf · 2026-07-17

This work discusses the issue of **statistical self-consistency** in large language models (LLMs). If prompts are understood as conditional distribution estimates, model outputs should satisfy basic probability identities, such as the law of total probability. The authors use a **binary tree hierarchical partitioning** to divide populations into subgroups, querying the model for subgroup estimates, aggregating them back to the overall population, and comparing this with direct overall estimates. Results show that frontier models generally violate these consistency properties. The paper also identifies a phenomenon termed the **macro fallacy**: - Overall estimates reconstructed from finer-grained subgroup answers are often closer to human reference data than direct overall estimates. - This effect persists across different tree structures and estimation tasks. - It can be partially mitigated through implicit prompting. The authors conclude that models actually possess relevant subgroup knowledge but fail to stably propagate this knowledge to macro-level aggregated judgments. Therefore, **statistical self-consistency** can serve as a reference-free evaluation metric.

Related event: Study Reveals LLMs Lack Statistical Self-Consistency(2 posts)→

Original post →

More from Research

Research channel →