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

A new benchmark reveals that leading vision-language models struggle with basic visual temporal reasoning, a skill humans perform with high accuracy. The evaluation, called CHRONOSIGHT, exposes a performance gap its authors term “chronological blindness” [1]. The CHRONOSIGHT benchmark tests models across five dimensions of temporal understanding: CHRONORANK, CHRONOLOCATE, CHRONODELTA, CHRONOREVERSE, and CHRONOODD [1]. It comprises 1,000 items drawn from eight process families, including biological growth, construction, and astronomical phenomena, with timescales ranging from minutes to millennia [1]. Human participants averaged 0.89 across the tasks, while the best-performing open-source model, Qwen2.5-VL-7B, scored 0.40 under direct prompting [1]. Several other evaluated models failed entirely on certain tasks due to an inability to format structured outputs, recording 100 percent parse-failure rates that masked their representational capabilities [2]. The difficulty models face with temporal order is not unique to CHRONOSIGHT. A separate study introducing a chronological reasoning benchmark found that vision-language models frequently exploit superficial cues, such as treating grayscale images as old and color images as new, rather than performing genuine chronological analysis [3]. That work noted that when these visual shortcuts conflict with chronological facts, model reasoning often collapses [3]. Another evaluation framework, TIME10k, assessed 37 vision-language models and found that temporal information is structured along a low-dimensional, non-linear manifold in the embedding space, suggesting that explicit timeline representations can improve temporal reasoning accuracy [4]. The CHRONOSIGHT authors probed whether the performance gap stems from visual perception deficits or instruction-following limitations. Using lightweight LoRA fine-tuning on only 151 training examples, they raised CHRONODELTA accuracy from near-zero to 0.43 [1]. The adapter transferred zero-shot to related tasks, lifting CHRONOREVERSE accuracy to 0.64 and CHRONOODD to 0.37, indicating that the bottleneck is partially one of instruction following rather than visual perception [1]. The benchmark, evaluation code, and raw model predictions will be released upon acceptance to facilitate reproducible follow-up research [2].