Research map / Large language models / Language, safety and interpretability
Parse Trees Guided LLM Prompt Compression
Wenhao Mao, Chengbin Hou, Tianyu Zhang, Xinyu Lin, Ke Tang, Hairong Lv
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add first node adjustment (a2) in value adjustment (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 70.59 -> 74.57 | Improved |
| add hierarchical global tree with node value adjustment over variant 1 (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 70.79 -> 74.57 | Improved |
| add global parse tree without value adjustment on top of local trees (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 71.06 -> 70.79 | Hurt |
| add local parse trees to entropy-only compression (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 69.75 -> 71.06 | Improved |
| add information entropy to parse-tree-only compression (global tree) (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 70.44 -> 70.79 | Improved |
| add information entropy to parse-tree-only compression (local trees) (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 69.33 -> 71.06 | Improved |
| add local entropy approximation to the full method (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 74.04 -> 74.57 | Improved |
| add local entropy approximation to global tree without adjustment (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 71.37 -> 70.79 | Hurt |
| add local entropy approximation to local-tree variant (ablation) | PartPrompt · BBCnews (50% tokens) | BS-F1: 70.29 -> 71.06 | Improved |
| use PartPrompt instead of LLMLingua2 on Llama3-70B | PartPrompt · BBCnews (50% tokens), Llama3-70B | BLEU: 13.60 -> 25.87 | Improved |
| use PartPrompt instead of LLMLingua on GSM8K | PartPrompt · GSM8K (30% tokens) | EM: 67.85 -> 53.30 | Mixed |
| use PartPrompt on RULER-niah instead of the uncompressed prompt | PartPrompt · RULER niah | score: 21.0 -> 69.0 | Improved |
Closest papers by meaning
- Fundamental Limits of Prompt Compression: A Rate-Distortion Framework for Black-Box Language Models (NeurIPS 2024 · similarity 0.88)
- Leveraging Attention to Effectively Compress Prompts for Long-Context LLMs (AAAI 2025 · similarity 0.88)
- Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.88)
- Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference (AAAI 2025 · similarity 0.88)
- LLMCBench: Benchmarking Large Language Model Compression for Efficient Deployment (NeurIPS 2024 · similarity 0.86)
- AdmTree: Compressing Lengthy Context with Adaptive Semantic Trees (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
