Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models
LLMs can know the parts and still miss the whole.
The paper tests whether model estimates obey the law of total probability when a population is split into subgroups and recombined. Across domains and frontier models, the authors report widespread failures of that consistency check. In persona prompting, finer subgroup estimates often match human reference data better than direct population-level answers, a pattern they call the macro fallacy. They frame statistical self-consistency as a reference-free evaluation criterion for LLMs. HF Daily Papers' note
The paper tests whether model estimates obey the law of total probability when a population is split into subgroups and recombined. Across domains and frontier models, the authors report widespread failures of that consistency check. In persona prompting, finer subgroup estimates often match human reference data better than direct population-level answers, a pattern they call the macro fallacy. They frame statistical self-consistency as a reference-free evaluation criterion for LLMs. HF Daily Papers' note
score 4