When Time Becomes Sparse: Evaluating Temporal Reasoning Robustness in Large Language Models via Anchoring Density Control

Abstract

Temporal reasoning is fundamental to information processing systems that organize and interpret event-centric knowledge. Existing evaluations, however, typically assess large language models (LLMs) in contexts densely grounded by explicit timestamps, where strong temporal anchors simplify chronological inference. We introduce DLGD<inline-formula> <tex-math notation="LaTeX">${}^{\prime }$ </tex-math></inline-formula>(Density-Lowered Grounding Degradation), a controlled robustness framework that systematically reduces temporal anchoring density while maintaining the gold chronological order and aiming to preserve event semantics. Using DLGD<inline-formula> <tex-math notation="LaTeX">${}^{\prime }$ </tex-math></inline-formula>, we construct hybrid-time narratives by partially replacing explicit timestamps with relative temporal expressions and evaluate temporal reasoning as global event ordering. For models with non-trivial event-ordering ability, performance consistently degrades under reduced anchoring. For example, QwQ-32B drops from 0.54 to 0.33 in Exact Match and from 0.73 to 0.53 in Kendall’s <inline-formula> <tex-math notation="LaTeX">$\tau $ </tex-math></inline-formula> from Explicit-Time (ET) to Density-Reduced (DR), while pairwise consistency declines only from 0.93 to 0.88. Masking a single temporal anchor produces a descriptive Exact Match reduction of up to 26.7% under ET, and temporal judgment reveals substantial confidence miscalibration, with some models rejecting up to 86.5% of correct orderings. These results show that LLM temporal reasoning is highly sensitive to the availability and distribution of explicit temporal anchors, particularly for global sequence reconstruction. This sensitivity motivates further evaluation across broader document domains, discourse-level temporal dependencies, and more ambiguous or adaptive temporal reasoning settings.