What Really Controls Temporal Reasoning in Large Language Models: Tokenisation or Representation of Time?
This work addresses temporal reasoning challenges in LLMs for multilingual and multi-calendar applications, providing insights into resource-dependent bottlenecks, but it is incremental as it builds on existing benchmarking and analysis methods.
The study tackled the problem of temporal reasoning in large language models by evaluating 20 LLMs on MultiTempBench, a multilingual benchmark with 15,000 examples across five languages and multiple calendars, finding that tokenisation quality is a bottleneck in low-resource settings where accuracy collapses, while temporal linearity is key in high-resource languages.
We present MultiTempBench, a multilingual temporal reasoning benchmark spanning three tasks, date arithmetic, time zone conversion, and temporal relation extraction across five languages (English, German, Chinese, Arabic, and Hausa) and multiple calendar conventions (Gregorian, Hijri, and Chinese Lunar). MultiTempBench contains $15,000$ examples built by translating $750$ curated English questions and expanding each into controlled date-format variants. We evaluate 20 LLMs and introduce the multilingual Date Fragmentation Ratio (mDFR), calibrated with human severity ratings, together with geometric-probing analyses of internal temporal representations. We find tokenisation quality of temporal artefacts is a resource-dependent bottleneck: in low-resource languages and rarer calendar formats, fragmentation disrupts Year/Month/Day separation and accuracy collapses, while high-resource settings are often robust to digit-level splitting. Beyond tokenisation, crossed mixed-effects regression shows that temporal linearity is the strongest predictor of temporal reasoning in high-resource languages, whereas fragmentation is the stronger predictor in low-resource languages. Code is available at: https://github.com/gagan3012/mtb