Tag
This paper identifies a curse of ambiguity in language models, where more ambiguous next-token distributions are harder to learn, tracing this to architectural and learning roots and validating on synthetic and real data.
This paper proposes a method to extend factorized probability distributions defined on non-identical variable sets to a common measurable space, enabling principled comparison using distributional discrepancy measures.
NoiseLang is a probabilistic programming language where every value is a distribution. It compiles to efficient Monte Carlo simulations using a JIT compiler and supports conditional inference.