Chronological Blindness: Benchmarking Temporal Reasoning in Vision-Language Models with CHRONOSIGHT

Human perception of visual scenes is inherently temporal. We instinctively recognize whether a fruit is ripening or rotting, whether construction is progressing or being demolished, and approximately how much time separates two photographs of the same subject. Whether large vision-language models (VLMs) share this competence remains an open and practically important question. We introduce CHRONOSIGHT, a rigorously controlled benchmark designed to evaluate five complementary dimensions of visual temporal reasoning: CHRONORANK (chronological ordering of image sequences), CHRONOLOCATE (ordinal stage localization from a single image), CHRONODELTA (estimation of the time elapsed between two images on a logarithmic time scale), CHRONOREVERSE (detection of temporally reversed sequences), and CHRONOODD (identification of a temporal outlier within a set). The benchmark comprises 1,000 items balanced across eight distinct process families (biological growth, food transformation, physical weathering, construction, environmental change, human aging, astronomical phenomena, and urban dynamics) spanning timescales from minutes to millennia. We evaluate eight open-source VLMs ranging from 500M to 19B parameters under two prompting regimes (direct, temporal-cue) and collect human performance baselines. Our principal findings are three-fold. First, there is a substantial and consistent gap between human performance (mean ≈ 0.89 across tasks) and the best open model evaluated (Qwen2.5-VL-7B, 0.40 under direct prompting), a gap we term chronological blindness. Second, models differ dramatically in their ability to format structured outputs; several models with 100% parse-failure rates score zero on some tasks, masking their true representational capabilities. Third, lightweight supervised fine-tuning via low-rank adaptation (LoRA) ononly151trainingexamplesraisesCHRONODELTA accuracy from near-zero to 0.43, and this adapter transfers zero-shot to related tasks (CHRONOODD: 0.37; CHRONOREVERSE: 0.64), suggesting that the bottleneck is partially one of instruction following rather than visual perception. The benchmark, evaluation code, and all raw model predictions will be released upon acceptance to facilitate reproducible follow-up research.