Analyzing constrained LLM through PDFA-learning

TL;DR

This paper introduces a novel congruence-based method for analyzing constrained language models, enabling efficient learning of their statistical properties during text generation.

Contribution

It develops an algorithm to learn the quotient with respect to a new congruence that handles null next-symbol probabilities in constrained LLMs.

Findings

Effective analysis of statistical properties of constrained LLMs

Algorithm demonstrates efficiency in learning the quotient

Applicable to various case studies of language model constraints

Abstract

We define a congruence that copes with null next-symbol probabilities that arise when the output of a language model is constrained by some means during text generation. We develop an algorithm for efficiently learning the quotient with respect to this congruence and evaluate it on case studies for analyzing statistical properties of LLM.