LLM Performance Drops in 'Text Uncanny Valley' with Word Fragmentation

A recent investigation published on arXiv (2605.07186) indicates that Large Language Models (LLMs) experience a non-linear decline in performance when encountering text with corrupted word boundaries. Researchers discovered that by adding whitespace within words, the accuracy of detection follows a U-shaped trajectory as the insertion rate increases, a phenomenon referred to as the 'Text Uncanny Valley.' The mode transition hypothesis proposed suggests that LLMs transition from processing at the word level to the character level, with the valley illustrating a chaotic shift where neither processing mode is effective. Four experiments and one analysis corroborate this, revealing that in-context learning does not enhance performance at the valley's lowest point, while regularizing the perturbation mitigates the U-shape. Similar trends were noted in a mathematical reasoning task.