AI★★★arXiv · 2026-07-16
Statistical Self-Consistency in Language Models through Partition, Prompt, Aggregate
This paper investigates whether large language models adhere to statistical self-consistency principles during in-context learning, particularly how prior-weighted conditional distributions aggregate into population-level marginals.
📌 Key points
- LLM estimates should satisfy basic probabilistic identities
- Focus on how prior-weighted conditional distributions aggregate
- Validate LLM consistency through partitioning
本页为 gitzw.com 基于公开来源的 AI 中文解读，非原文转载。