Brief

Researchers investigated temporal concept drift in legal judgment prediction by training transformer models on Ukrainian court decisions from different geopolitical eras. They found that models trained on older data performed significantly worse on newer data, indicating a severe forward degradation in predictive accuracy. While legal-domain pretraining offered some mitigation, chronological continual learning proved effective in preventing catastrophic forgetting and improving performance over time. The study highlights that legal language evolution, influenced by geopolitical events, is additive and presents a significant challenge for AI models. AI

Researchers have developed a new benchmark, UA-StatuteRetrieval, to assess the stability of co-citation predictability in legal information systems over time. Analyzing 396 million Ukrainian court citations from 2007 to 2026, they found a significant decay in retrieval performance, with predictability dropping by up to 47%. While high-frequency articles and criminal procedure maintained stability, mid-frequency articles and civil law showed notable degradation, partly explained by a 2017 judicial reform and a 4.3% semantic shift in article citation patterns. AI