Research map / Large language models / Language, safety and interpretability
From Token to Token Pair: Efficient Prompt Compression for Large Language Models in Clinical Prediction
Mingcheng Zhu, Zhiyao Luo, Yu Liu, Tingting Zhu
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 self-supervised fine-tuning of new TPE embeddings (ablation) | MedTPE (Llama3-1B) · MIMIC-IV ICU Mortality | F1: 0.000 -> 0.109 | Improved |
| add self-supervised fine-tuning of new TPE embeddings (ablation) | MedTPE (Qwen2.5-1.5B) · MIMIC-IV Phenotyping | F1: 0.064 -> 0.218 | Improved |
| use dependency-aware replacement instead of expanding vocabulary (ablation) | MedTPE (Llama3-1B) · MIMIC-IV Phenotyping | F1: 0.074 -> 0.114 | Improved |
| use dependency-aware replacement instead of expanding vocabulary (ablation) | MedTPE (Llama3-1B) · MIMIC-IV ICU Mortality | inference time (min): 79.2 -> 39.3 | Improved |
| replace LLMLingua2 token removal with MedTPE merging | Llama3-1B · MIMIC-IV Phenotyping | F1: 0.058 -> 0.114 | Improved |
| replace ZeTT hypernetwork embeddings with MedTPE | Llama3-1B · MIMIC-IV ICU Mortality | inference time (min): 201.8 -> 39.3 | Improved |
| replace T5Summary with MedTPE | Qwen2.5-1.5B · MIMIC-IV Phenotyping | F1: 0.073 -> 0.218 | Improved |
| apply MedTPE to Qwen2.5-1.5B on ICU mortality | MedTPE · MIMIC-IV ICU Mortality | F1: 0.122 -> 0.122 | No clear effect |
| add embedding fine-tuning to LLMLingua2 baseline (ablation) | LLMLingua2 (Llama3-1B) · MIMIC-IV Phenotyping | F1: 0.058 -> 0.017 | Hurt |
| use MIMIC-IV-trained tokeniser instead of in-domain on EHRSHOT (ablation) | MedTPE (Llama3-1B) · EHRSHOT 30-day Readmission | F1: 0.191 -> 0.014 | Hurt |
| increase replaced-token budget beyond 5,000 (ablation) | MedTPE · MIMIC-IV and EHRSHOT | compression rate | No clear effect |
| use bigrams (N=2) versus larger n-grams (ablation) | MedTPE · MIMIC-IV and EHRSHOT | compression rate | Improved |
Closest papers by meaning
- Multimodal Medical Code Tokenizer (ICML 2025 · similarity 0.88)
- Rep2Text: Decoding Full Text from a Single LLM Token Representation (NeurIPS 2026 · similarity 0.87)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.87)
- One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models (NeurIPS 2026 · similarity 0.84)
- zip2zip: Inference-Time Adaptive Tokenization via Online Compression (NeurIPS 2025 · similarity 0.84)
- MedAlign: A Clinician-Generated Dataset for Instruction Following with Electronic Medical Records (AAAI 2024 · similarity 0.84)
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.
