LLMs Compress 10 Intensity Words into 5 Distinct Outputs

A study on arXiv (2605.21827) investigates whether language models preserve ordinal meaning of intensity words when producing numeric actions. Using a researcher-constructed scale of 10 English degree modifiers from slightly to drastically, informed by Quirk et al.'s taxonomy, Claude Haiku received natural-language instructions in a controlled resource-allocation environment. The only variable changed between runs was the intensity word or starting system state. Across 6,620 runs at temperatures 0.0 and 0.7, three patterns emerged: the model compressed 10 words into 5 distinct median outputs, with four lower-tier words mapping to the same value (Spearman rho = 0.845, p < 0.001); current system state influenced output; and higher temperatures increased variability.