Research map / Large language models / Language, safety and interpretability
STT-LLM: Structural-Temporal Tokenization for Adapting LLMs to Longitudinal Clinical Profiles
Maxx Richard Rahman, Mostafa Hammouda, Wolfgang Maass
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 |
|---|---|---|---|
| use structural-temporal tokenization instead of native text tokenization | LLaMA-3.1 (8B) · Steroid-M zero-shot global anomaly detection | F1: 0.19 -> 0.26 | Improved |
| use STT-LLM tokenization across all seven backbone LLMs | seven 7-8B LLMs · Steroid-M | sensitivity / RMSE: +2 to +14 pp sensitivity; RMSE reductions 9.35% to 26.26% | Improved |
| add structural tokenizer (ablation) | STT-LLM · Steroid-M | sensitivity (global anomaly): 0.0968 -> 0.1935 | Improved |
| add temporal tokenizer (ablation) | STT-LLM · Steroid-M | sensitivity (global anomaly): 0.1237 -> 0.1935 | Improved |
| add embedding layer (ablation) | STT-LLM · Steroid-M | AUC: 0.5352 -> 0.5675 | Improved |
| keep all components versus removing all (ablation) | STT-LLM · Steroid-M | RMSE: 1687.71 -> 1664.59 | Improved |
| remove only some components (partial ablation) versus removing all (ablation) | STT-LLM · Steroid-M | global anomaly sensitivity: 0.1398 (w/o all) -> 0.0645 (w/o structural + temporal) | Hurt |
| increase in-context examples from 5 to 20 | baseline LLMs (native tokenization) · few-shot sequence prediction | RMSE/MAE | No clear effect |
| increase in-context examples from 2 to 20 | STT-LLM · Steroid-M few-shot global anomaly detection | F1: 0.15 -> 0.70 | Improved |
| STT-LLM versus specialized non-LLM anomaly detector M-GNN | STT-LLM · Steroid-M global anomaly detection | F1: 0.6292 (M-GNN) vs 0.2637 | Hurt |
Closest papers by meaning
- Data-Driven Discovery of Dynamical Systems in Pharmacology using Large Language Models (NeurIPS 2024 · similarity 0.84)
- Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction (NeurIPS 2026 · similarity 0.83)
- Enhancing Aging Biomarker Research through Large Language Models and Knowledge Graphs (Student Abstract) (AAAI 2025 · similarity 0.83)
- Tag-LLM: Repurposing General-Purpose LLMs for Specialized Domains (ICML 2024 · similarity 0.83)
- CGBench: Benchmarking Language Model Scientific Reasoning for Clinical Genetics Research (NeurIPS 2025 · similarity 0.83)
- Fragments to Facts: Partial-Information Fragment Inference from LLMs (ICML 2025 · similarity 0.83)
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.
