FIT: Defying Catastrophic Forgetting in Continual LLM Unlearning

We introduce FIT, a scalable framework for continual unlearning, together with the PCH benchmark for evaluating forgetting effectiveness, utility preservation, and robustness in large language models.

Xiaoyu Xu¹, Minxin Du¹*, Kun Fang¹, Zi Liang¹, Yaxin Xiao¹, Zhicong Huang², Cheng Hong², Qingqing Ye¹, Haibo Hu¹* ¹The Hong Kong Polytechnic University ²Ant Group * Corresponding author

Paper Model Dataset Code

600

PCH size

600

QA-Pairs size

3

Evaluation Criteria

2026

Release time

Introduction

The FIT benchmark evaluates whether large language models can reliably “forget” targeted information under continual unlearning. Unlike single-shot deletion, real-world requests arrive sequentially, often causing parameter drift, unstable gradients, and progressive utility degradation if handled with naïve unlearning strategies. FIT provides a principled framework to diagnose these challenges and assess whether a model can remain stable while removing sensitive knowledge.

The PCH benchmark includes unified synthetic corpora, well-defined evaluation protocols, and reproducible scripts. By testing models across long sequences of deletion steps, FIT offers a holistic view of forgetting effectiveness, utility preservation, and robustness against post-unlearning attacks—enabling deeper analysis of whether undesired knowledge is truly removed or merely obscured.

Citation

If you find our code and paper helpful, please consider citing our work:

@misc{xufit2026,
  title= {{FIT}: Defying Catastrophic Forgetting in Continual {LLM} Unlearning},
  author={Xiaoyu Xu and Minxin Du and Kun Fang and Zi Liang and Yaxin Xiao and Zhicong Huang and Cheng Hong and Qingqing Ye and Haibo Hu},
  howpublished={arXiv:2601.21682},    
  year={2026}}

FIT: Defying Catastrophic Forgetting in Continual LLM Unlearning

600

600

3

2026

Introduction

New Evaluation Criteria

Evaluation Corpora

Personal Information

Copyright

Harmful Content

Citation