A new dataset and testing harness called date-math-bench reveals that large language models struggle with basic date arithmetic. Across 101 randomized questions per model, common errors included miscalculating elapsed days versus ordinal day counts and incorrectly predicting the next day of the week. Specifically, business day calculations proved to be a significant blind spot for models like Claude Haiku, GPT-4o-mini, and Llama 3.3 70B, while Claude Sonnet and Qwen3-27B performed perfectly in this category. AI
IMPACT Highlights a critical gap in LLM reasoning capabilities, potentially impacting applications requiring temporal understanding.
RANK_REASON New benchmark and dataset released for evaluating LLM performance on date arithmetic. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
