Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 28·5 min read
Lost in Compression: A Controlled Cross-Lingual Audit of Extractive Prompt Compressors
TL;DR
Extractive prompt compression techniques, particularly those trained on English data, struggle with non-English languages, leading to significant context loss. A multilingual compressor, XProvence, shows promise in mitigating this issue, but aggressive settings can still result in poor performance.
✦ Why It Matters
Evaluate your compression strategies for non-English languages to avoid significant context loss.
Key Takeaways
- Assess the performance of learned compressors on non-English languages.
- Consider using translate-then-compress methods for better efficiency.
- Monitor context retention rates when using aggressive compression settings.
- Be cautious with compression budgets, especially outside English.
How It Works
Extractive prompt compression works by identifying and removing low-information tokens from input data to reduce the number of tokens processed by LLMs. The compressors are trained on specific language data, which influences their effectiveness across different languages.
The study compares learned compressors, which adapt based on training data, to deterministic methods that apply fixed rules for compression.