Researchers have developed a new method called ATLAS (Authorship Tracing via Likelihood-based Attribution of Sources) to improve the accuracy of identifying which large language model generated a given text. Existing methods often fail when encountering text from unseen domains or future model versions. ATLAS addresses this by modeling how text is generated from each source family and comparing likelihoods, rather than direct prediction. The effectiveness of ATLAS is demonstrated on MAABench, a new benchmark designed for out-of-distribution scenarios, where it significantly outperforms previous approaches, particularly when both domain and model version have changed. Additionally, ATLAS shows promise in identifying distillation sources by measuring how closely a target LLM resembles known teacher models. AI
IMPACT Enhances AI safety and academic integrity by providing a more robust method for detecting AI-generated text, even from unseen models.
RANK_REASON The cluster contains an academic paper detailing a new method for model authorship attribution. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
