LLM-Based Compression

Making ML-powered compression fast, robust, and practical for data systems.

Project Leads: Luisa Neubauer and Mihail Stoian

Description

LLM-Based Compression is a Data Systems Lab project at the University of Technology Nuremberg (UTN) exploring two lossless approaches: compressing text with LLM predictions and entropy coding, and using coding agents to create codecs for specific datasets.

We study when storage savings justify compression and decompression costs, how to improve throughput, and how LLM-based compression handles adversarial inputs. Synthesized codecs are verified by reconstructing data byte for byte, then run without model inference while targeting fast decompression and random access.

Collaborators

David Adams — Doctoral Researcher, University of Melbourne
Robin Sternberg — Student Researcher
Ivan Iachnyk — Student Researcher

Waiting to Decompress: The Economics of LLM-Based Compression

Andreas Kipf, Tobias Schmidt, Ping-Lin Kuo, Skander Krid, Moritz Rengert, Luca Heller, Andreas Zimmerer, Mihail Stoian, Varun Pandey, and Alexander van Renen

In 16th Conference on Innovative Data Systems Research, CIDR 2026, Chaminade, HI, USA, January 18-21, 2026

Bib PDF Code

@inproceedings{kipf2025llmcompress,
  author = {Kipf, Andreas and Schmidt, Tobias and Kuo, Ping-Lin and Krid, Skander and Rengert, Moritz and Heller, Luca and Zimmerer, Andreas and Stoian, Mihail and Pandey, Varun and van Renen, Alexander},
  title = {Waiting to Decompress: The Economics of {LLM}-Based Compression},
  booktitle = {16th Conference on Innovative Data Systems Research, {CIDR} 2026, Chaminade, HI, USA, January 18-21},
  publisher = {www.cidrdb.org},
  year = {2026},
}

High-Throughput LLM-Based Compression and What It Takes

David Adams, Luisa Neubauer, Ping-Lin Kuo, Mihail Stoian, Thomas Rückstieß, Josif Grabocka, Renata Borovica-Gajic, and Andreas Kipf

In Sixteenth International Workshop on Accelerating Analytics and Data Management Systems Using Modern Processor and Storage Architectures (ADMS), 2026

Bib PDF

@inproceedings{adams2026fastllmcompression,
  title = {High-Throughput {LLM}-Based Compression and What It Takes},
  author = {Adams, David and Neubauer, Luisa and Kuo, Ping-Lin and Stoian, Mihail and Rückstieß, Thomas and Grabocka, Josif and Borovica-Gajic, Renata and Kipf, Andreas},
  year = {2026},
  maintitle = {Proceedings of the {VLDB} Endowment},
  booktitle = {Sixteenth International Workshop on Accelerating Analytics and Data Management Systems Using Modern Processor and Storage Architectures (ADMS)},
  publisher = {{VLDB} Endowment},
}

When Adversarial Data Still Compresses: Towards Robust LLM-Based Compression

Luisa Neubauer, David Adams, Mihail Stoian, Andreas Zimmerer, Alexander van Renen, Renata Borovica-Gajic, and Andreas Kipf

In NeurIPS 2026 Workshop on Machine Learning for Systems (ML for Systems), 2026

Bib PDF

@inproceedings{neubauer2026adversarial,
  title = {When Adversarial Data Still Compresses: Towards Robust {LLM}-Based Compression},
  author = {Neubauer, Luisa and Adams, David and Stoian, Mihail and Zimmerer, Andreas and van Renen, Alexander and Borovica-Gajic, Renata and Kipf, Andreas},
  year = {2026},
  booktitle = {NeurIPS 2026 Workshop on Machine Learning for Systems ({ML} for Systems)},
}

The Case for Synthesized Data Compression

Ivan Iachnyk, Mihail Stoian, Luisa Neubauer, Ryan Marcus, and Andreas Kipf

CoRR, 2026

Bib PDF Code

@article{iachnyk2026synthesizedcompression,
  title = {The Case for Synthesized Data Compression},
  author = {Iachnyk, Ivan and Stoian, Mihail and Neubauer, Luisa and Marcus, Ryan and Kipf, Andreas},
  journal = {CoRR},
  volume = {abs/2610.09913},
  year = {2026},
  eprinttype = {arXiv},
  eprint = {2610.09913},
  archiveprefix = {arXiv},
  primaryclass = {cs.DB},
}

Description

Collaborators

Publications