LLMs show inconsistent belief updates from evidence, with learned heuristics sometimes beating exact Bayesian computation due to misspecified world models.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
LLMs are not (consistently) Bayesian: Quantifying internal (in)consistencies of LLMs' probabilistic beliefs
LLMs show inconsistent belief updates from evidence, with learned heuristics sometimes beating exact Bayesian computation due to misspecified world models.