Showing 1–1 of 1 results for author: Garat, A

Analyzing constrained LLM through PDFA-learning

Authors: Matías Carrasco, Franz Mayr, Sergio Yovine, Johny Kidd, Martín Iturbide, Juan Pedro da Silva, Alejo Garat

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. 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. △ Less

Submitted 15 June, 2024; v1 submitted 12 June, 2024; originally announced June 2024.

Comments: Workshop Paper