Lexical Prompt Compression
A training-free, auxiliary-model-free prompt compression approach introduced in an arXiv paper. It applies eleven composable lexical-level NLP transformations to compress prompts, requires only CPU resources, and produces fully deterministic compression results.
Timeline (last 90 days)
The arXiv paper proposes a pure CPU prompt compression pipeline that does not rely on training, does not rely on any auxiliary models, and uses only lexical-level NLP processing
The compression pipeline consists of eleven independently switchable lexical transformations, including stopword removal, filler phrase deletion, abbreviation and contraction replacement, POS-based pruning, lemmatization, WordNet-based synonym shortening, and named entity preservation
The most aggressive configuration achieves an average token reduction rate of 40.3% (standard deviation 9.2), with a BERTScore-F1 of 0.876 relative to the output of the original prompt
The stopword-removal-only configuration can achieve a token reduction of 29.6% while maintaining a BERTScore-F1 of 0.913
Under aggressive compression configurations, commonsense reasoning tasks exhibit systematic failure patterns
Different tasks have different tolerances for prompt compression, with factual question-answering or structured tasks tolerating compression better than reasoning-intensive tasks
The named entity preservation mechanism explicitly protects core information components such as person names, place names, and proper nouns during the compression process
There exists a Pareto frontier trade-off between compression strength and output quality, where higher compression ratios require accepting a certain degree of quality degradation.
The authors have publicly released all code, prompts, and per-cell results to support reproducibility.
All Facts (9)
The arXiv paper proposes a pure CPU prompt compression pipeline that does not rely on training, does not rely on any auxiliary models, and uses only lexical-level NLP processing
50%UnverifiedThe compression pipeline consists of eleven independently switchable lexical transformations, including stopword removal, filler phrase deletion, abbreviation and contraction replacement, POS-based pruning, lemmatization, WordNet-based synonym shortening, and named entity preservation
50%UnverifiedThe most aggressive configuration achieves an average token reduction rate of 40.3% (standard deviation 9.2), with a BERTScore-F1 of 0.876 relative to the output of the original prompt
50%UnverifiedThe stopword-removal-only configuration can achieve a token reduction of 29.6% while maintaining a BERTScore-F1 of 0.913
50%UnverifiedUnder aggressive compression configurations, commonsense reasoning tasks exhibit systematic failure patterns
50%UnverifiedDifferent tasks have different tolerances for prompt compression, with factual question-answering or structured tasks tolerating compression better than reasoning-intensive tasks
50%UnverifiedThe named entity preservation mechanism explicitly protects core information components such as person names, place names, and proper nouns during the compression process
50%UnverifiedThere exists a Pareto frontier trade-off between compression strength and output quality, where higher compression ratios require accepting a certain degree of quality degradation.
50%UnverifiedThe authors have publicly released all code, prompts, and per-cell results to support reproducibility.
50%