Temporal-LLM-KGC: A Diagnostic Study of Temporal Evidence in Large Language Model Knowledge Graph Completion

Large language models (LLMs) have shown promise for temporal knowledge graph completion (temporal KGC), but accuracy alone does not reveal dependence on historical context. We evaluate Normal, No History, Anonymous, and Noisy evidence conditions on ICEWS14 and ICEWS18 with four instruction-tuned LLMs. We also test Temporal Evidence Filtering (TEF), a deterministic, training-free proof-of-concept selector combining relation compatibility, temporal recency, and entity-role consistency under a fixed ten-fact prompt budget. Removing or displacing history lowers accuracy; Anonymous improves it but simultaneously changes the constrained candidate vocabulary, precluding a single-mechanism explanation. TEF has positive descriptive differences in all 24 dataset–model–metric contrasts. Across eight dataset–model pairs, its equal-weight macro Hits@10 gain is +0.031375 (nominal paired-bootstrap 95% confidence interval: [+0.021875, +0.040875]); 10 of the 24 individual contrasts remain significant after Holm correction. Qwen3-4B selector baselines, component ablations, parameter sensitivity, and a disjoint second query sample qualify this result. TEF adds a mean 4.360 and 19.071 ms of CPU selection per query on ICEWS14 and ICEWS18, respectively; it does not compress the Normal prompt or establish an end-to-end speedup. Conclusions concern tail prediction with constrained generated lists on these two datasets, not exhaustive ranking or an identified internal reasoning mechanism.