LLMs fail when facts change — and current fixes don't work in real time

A new benchmark tested how well large language models adapt when the world changes — facts shift, events unfold, entities evolve over months or years. It turns out current methods, including the most popular approach (retrieval-augmented generation, which feeds models fresh information), struggle badly: they forget old knowledge, contradict themselves across time, and can't track how situations actually develop.

Every deployed language model learns from a snapshot of the world that gets older the moment it ships. The field has three main approaches to fix this: retraining the model with new data, editing specific facts into memory, or feeding fresh information at question time. This paper shows all three fail in settings that matter — real events that unfold over months with cascading dependencies, where you need to understand not just what changed but how and when. It exposes a gap between what these systems can do in lab benchmarks and what they actually do when used for anything that requires temporal consistency. This matters because companies are shipping these models into production for legal research, medical updates, financial analysis, and news, all domains where temporal inconsistency isn't a curiosity — it's a liability.