LLMs Are Not (Consistently) Bayesian: Quantifying Internal (In)consistencies of LLMs’ Probabilistic Beliefs
2026-08-28 08:00Models🔥 28.9 heat score
1sources
1days unfolding
28.9heat score
0mentions
SummaryAI generated
The Apple Machine Learning Team released a research report stating that large language models (LLMs) exhibit internal inconsistencies in their probabilistic belief updates when dealing with complex tasks without a single correct answer. The study considers LLMs as systems that follow specific information processing rules and uses the “Bayesian update gap” method to evaluate their performance in key fields such as medicine, science, and law. Experimental results show that existing LLMs fail to meet the consistency standard in rationally updating uncertain beliefs based on new evidence, which poses challenges for complex scenarios where AI is used for decision-making.
When modern AI systems are deployed in complex fields such as medicine, science, and law, they often face situations where there is no single correct answer. It is necessary to be able to update uncertain beliefs based on new evidence to make rational decisions. Researchers have introduced a technique that treats large language models as information processing rules, using “Bayesian update gap” to study how LLMs update the internal consistency of probability beliefs based on evidence. Existing experiments have widely evaluated this method.