LANTERN in the Event Stream: Training-Free Temporal Knowledge Graph Forecasting by Balancing Inertia and Shifts

Temporal knowledge graph forecasting (TKGF) asks a model to rank the most plausible future entity for a query such as (s, r, ?, t) from historical events. Recent training-free methods use large language models (LLMs) for this task, but their accuracy depends heavily on which past events are shown in the prompt under a tight context budget. We present LANTERN, a training-free prompting framework that addresses this bottleneck by combining two complementary views of history: a long-window strength score for stable interaction patterns and a short-window novelty score for sudden changes. LANTERN first filters unhelpful events, then selects a compact evidence set with Pareto-greedy selection, and finally adds one structure-aware analogical demonstration. Across ICEWS14, ICEWS05-15, ICEWS18, and GDELT, LANTERN consistently outperforms the state-of-the-art training-free baseline AnRe under the same backbone and 2-hop candidate protocol, improving Hits@1 by up to 2.5 points and MRR by up to 1.2 points.