Skip to content

upunaprosk/Awesome-LLM-Compression-Safety

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

19 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Awesome LLM Compression Safety

What breaks when we compress large language models?

Model compression techniques (quantization, pruning, distillation, low-rank adaptation) often preserve benchmark accuracy, but can silently degrade fairness, robustness, calibration, and safety.

This is a curated list of research studying undesired effects of model compression in LLMs, VLMs, and multimodal models, with a focus on fairness, robustness, calibration, and safety.
Contributions are welcome!

Unlike existing efficiency or compression lists, this repository focuses on fairness-, robustness-, calibration-, and safety-related regressions rather than throughput or accuracy alone.

LLM Compression Safety


🆕 Recent Papers (2025)

  • Fair-GPTQ: Bias-Aware Quantization for Large Language Models (2025)
  • Understanding the Unfairness in Network Quantization (ICML 2025)
  • Decomposed Trust: Exploring Privacy, Adversarial Robustness, Fairness, and Ethics of Low-Rank LLMs (2025)
  • Compressed but Compromised? Jailbreaking in Compressed LLMs (2025)

Surveys


Scope

This list focuses on how compression methods (quantization, pruning, distillation, low-rank methods) affect:

  • fairness and bias
  • robustness and reliability
  • calibration and confidence
  • toxicity, alignment, and safety
  • faithfulness and trustworthiness

Papers that focus only on efficiency or aggregate accuracy, without analyzing behavioral, fairness, robustness, or safety effects, are out of scope.


Contents


Fairness & Bias

(newest first)


Robustness & Reliability

(newest first)


Calibration & Confidence

(newest first)


Toxicity & Safety

(newest first)


Contributing

Contributions are welcome.

  • Papers must study undesired effects of compression
  • Use arXiv abstract links when available
  • List the final venue if accepted
  • Keep entries concise and consistent
  • Papers may appear in multiple sections

Formatting example

- **Paper Title**  
  Author et al., Venue Year  
  https://arxiv.org/abs/XXXX.XXXXX

About

A curated list of papers, docs, and code on the undesired effects of model compression, including impacts on fairness, robustness, calibration, and toxicity. The project is continuously updated. Welcome to PR the works (papers, repositories) that are missed by the repo!

Topics

Resources

Stars

Watchers

Forks

Contributors