Analyzing constrained LLM through PDFA-learning
Formal Languages and Automata Theory
2024-06-18 v2 Artificial Intelligence
Machine Learning
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.
Cite
@article{arxiv.2406.08269,
title = {Analyzing constrained LLM through PDFA-learning},
author = {Matías Carrasco and Franz Mayr and Sergio Yovine and Johny Kidd and Martín Iturbide and Juan Pedro da Silva and Alejo Garat},
journal= {arXiv preprint arXiv:2406.08269},
year = {2024}
}
Comments
Workshop Paper