Time-Decay-Aware Reasoning: Maintaining Large Language Model Knowledge Freshness Without Full Retraining

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models store factual knowledge in their parameters, but this knowledge is fixed at the training cut-off point and becomes outdated as the world changes; therefore, although fluent, the answers may be incorrect. A time-decay-aware reasoning framework has been introduced in this paper to maintain such knowledge over time without retraining or modifying the model’s parameters. The framework assigns a decay rate to each fact based on how often it is updated in a temporal knowledge base, converts this rate into the probability that a parametric answer has expired, and only issues a retrieval call when this probability exceeds a threshold; thus, retrieval is focused on facts that have actually aged. On TempLAMA (50,310 temporally scoped Wikidata questions, 2010–2020), the method outperforms the fine-tuning and editing baselines, recovers about two-thirds of the accuracy gap between a frozen model and exhaustive retrieval while issuing only half as many retrieval calls; it does not modify static knowledge. The Design aims to run large language models efficiently in the face of continuous fact updates.