READER: Robust Evidence-based Authorship Decoding via Extracted Representations

READER (Robust Evidence-based Authorship Decoding via Extracted Representations) is a new framework designed to identify the source of text generated by various LLMs using dynamic black-box provenance techniques. It utilizes a frozen proxy LLM to analyze black-box outputs, employing Bayesian Evidence Accumulation to enhance accuracy, achieving top-1 accuracy of 31.0%-42.4% from a single response and 70.0%-84.0% from 50 responses on the Agent500 dataset. This advancement is significant for practitioners as it enables reliable attribution of text outputs to specific models, enhancing transparency in LLM applications.