Analyzing constrained LLM through PDFA-learning

• DOI: 10.48550/arxiv.2406.08269
• Publication Date: 2024-06-12
• Authors: Matías Carrasco; Franz Mayr; Sergio Yovine; Johny Kidd; Martín Iturbide; Juan Pedro da Silva; Alejo Garat
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Formal Languages and Automata Theory (cs.FL); Computer Science - Formal Languages and Automata Theory; Machine Learning (cs.LG)

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