What Really Controls Temporal Reasoning in Large Language Models: Tokenisation or Representation of Time?
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866910060157861888 |
|---|---|
| author | Bhatia, Gagan Isa, Ahmad Muhammad Peyrard, Maxime Zhao, Wei |
| author_facet | Bhatia, Gagan Isa, Ahmad Muhammad Peyrard, Maxime Zhao, Wei |
| contents | 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 |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_19017 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | What Really Controls Temporal Reasoning in Large Language Models: Tokenisation or Representation of Time? Bhatia, Gagan Isa, Ahmad Muhammad Peyrard, Maxime Zhao, Wei Computation and Language Artificial Intelligence 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 |
| title | What Really Controls Temporal Reasoning in Large Language Models: Tokenisation or Representation of Time? |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2603.19017 |