Probabilistic Attribution For Large Language Models

In three linesNovel probabilistic token attribution method for LLMs using conditional probabilities and Bayes rule to invert next-token log-probabilities. Captures model's internal representation of token sequence distribution independent of computational structure. Evaluates 8 models across 7 prompts to analyze token sensitivity, response stability, and training convergence.

Summary generated by Claude — human-verified