Analyzing constrained LLM through PDFA-learning
This addresses a technical bottleneck in analyzing constrained LLMs, but appears incremental as it builds on existing PDFA-learning methods.
The paper tackles the problem of analyzing constrained language models by defining a congruence to handle null next-symbol probabilities during constrained text generation, and develops an efficient algorithm for learning the quotient, evaluating it on case studies for statistical analysis of LLMs.
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.