Researchers have built a training-free prompt compressor that runs on a CPU and uses conventional text-processing rules instead of another language model. Its most aggressive configuration reduced prompt length by an average of 40.3%, with a standard deviation of 9.2 percentage points.
The configurable pipeline combines 11 transformations, including stopword and filler removal, shorter contractions and abbreviations, part-of-speech pruning, lemmatization, synonym shortening and protection for named entities. Because the rules are deterministic, the same prompt produces the same compressed result without auxiliary model inference.
The team evaluated 15 configurations on 1,242 English prompts drawn from six datasets and 11 automatically derived task categories. That produced 18,630 paired completions from GPT-4o-mini. The aggressive setup reached a BERTScore-F1 of 0.876 when its outputs were compared with answers to the original prompts. A stopword-only setup cut 29.6% of tokens and scored 0.913.
Compression was not equally safe for every use. Common-sense reasoning was a systematic failure mode under aggressive editing, showing that fewer tokens do not automatically preserve intent. The code, prompts and detailed results are released for reproduction. As a new arXiv preprint, the work still needs independent validation, including tests on models and languages beyond its English-only experiment.