Research map / Large language models / Instruction and fine-tuning
Privately Fine-Tuned LLMs Preserve Temporal Dynamics in Tabular Data
Lucas Rosenblatt, Peihan Liu, Ryan McKenna, Natalia Ponomareva
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 |
|---|---|---|---|
| replace flattened marginal synthesizer (AIM) with autoregressive PATH 4B | AIM · MIMIC-IV (eps=2) | TDCR (JSD): 0.736 -> 0.289 | Improved |
| replace AIM with PATH on univariate marginal divergence | AIM · MIMIC-IV (eps=2) | marginal divergence: 9.71 -> 3.30 | Improved |
| replace AIM with PATH 4B on HMM likelihood | AIM · Synthetic HMM (eps=10) | HMM likelihood (Wasserstein): 319.43 -> 50.99 | Improved |
| use AIM instead of PATH 4B for univariate marginals on synthetic HMM | PATH 4B · Synthetic HMM (eps=10) | marginal divergence: 0.83 -> 0.28 | Hurt |
| use PATH 4B instead of AIM on state transitions | AIM · MIMIC-IV | state transition Frobenius: 0.62 -> 0.38 | Improved |
| private fine-tuning (PATH 4B, eps=10) instead of Direct mechanism | Direct · MIMIC-IV (eps=10) | MAUVE: 0.596 -> 0.651 | Improved |
| private fine-tuning instead of non-private 10-shot Gemini 2.5 FL | Gemini 2.5 FL 10-shot · MIMIC-IV | MAUVE: 0.269 -> 0.651 | Improved |
| scale PATH from Gemma 1B to 4B at low privacy budget (ablation) | PATH · MIMIC-IV (eps=0.5) | MAUVE: 0.531 -> 0.627 | Improved |
| scale PATH from Gemma 1B to 4B at low privacy budget (ablation) | PATH · MIMIC-IV (eps=0.5) | marginal divergence: 3.93 -> 3.00 | Improved |
| scale PATH from Gemma 1B to 4B at eps=10 (ablation) | PATH · MIMIC-IV (eps=10) | MAUVE: 0.552 -> 0.651 | Improved |
| scale PATH from Gemma 1B to 4B on NYC 311 (eps=2) (ablation) | PATH · NYC 311 (eps=2) | MAUVE: 0.311 -> 0.507 | Improved |
| scale PATH from Gemma 1B to 4B on synthetic HMM (eps=2) (ablation) | PATH · Synthetic HMM (eps=2) | MAUVE: 0.307 -> 0.612 | Improved |
Closest papers by meaning
- Private Evolution Converges (NeurIPS 2025 · similarity 0.87)
- Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning (NeurIPS 2026 · similarity 0.86)
- Struct-Bench: A Benchmark for Differentially Private Structured Text Generation (NeurIPS 2025 · similarity 0.86)
- HARMONIC: Harnessing LLMs for Tabular Data Synthesis and Privacy Protection (NeurIPS 2024 · similarity 0.86)
- CuTS: Customizable Tabular Synthetic Data Generation (ICML 2024 · similarity 0.86)
- TabNAT: A Continuous-Discrete Joint Generative Framework for Tabular Data (ICML 2025 · similarity 0.85)
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.
