Effective Cutoff (LLM)

The effective cutoff of a large language model is the date at which its behavior on a specific topic stops reflecting current information. It is distinct from the reported knowledge cutoff announced by the model provider, and it can differ substantially across subjects within the same model. The term was introduced by Cheng, Marone, Weller, Lawrie, Khashabi, and Van Durme of Johns Hopkins in the 2024 paper "Dated Data: Tracing Knowledge Cutoffs in Large Language Models." They documented two main causes for the gap. First, temporal biases in CommonCrawl: each new crawl re-ingests large amounts of older content, so the training distribution skews toward the past even when the crawl itself is recent. Second, deduplication schemes that treat near-duplicates as a single document often retain an older revision and discard newer ones. Analyses of open-source models like Pythia and Llama showed effective cutoffs aligning to dates well before the reported cutoffs. The practical consequence is that a model with a reported 2024 cutoff may answer tax-code, regulatory, or library-API questions as if the year were 2022, and it may answer questions about long-tail topics as if the year were earlier still. The effective cutoff is one structural reason behind Date and Time Confusion in LLMs: when asked what year it is, models often guess a year consistent with their densest training signal rather than their nominal cutoff.