ZJUKLAB at SemEval-2025 Task 4: Unlearning via Model Merging

2025-03-27

AI Summary

To address the challenge of selective forgetting of sensitive knowledge in large language models—where conventional approaches often suffer from either over-forgetting or under-forgetting—this paper proposes a model-merging-based unlearning method grounded in TIES-Merging. First, sensitive knowledge is localized, and two specialized submodels (forgetting and retaining) are constructed; then, TIES-Merging is applied in weight space to achieve fine-grained, balanced forgetting control. Innovatively, this work introduces model merging into unlearning tasks and highlights the limitations of mainstream membership inference attacks (MIA) and ROUGE metrics for fine-grained forgetting evaluation. To enhance interpretability, the method integrates local loss analysis and weight-space diagnostics. Evaluated on SemEval-2025 Task 4, it achieves second place (Task Aggregate score: 0.944). The code is open-sourced, and comprehensive ablation and attribution analyses are provided.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

This paper presents the ZJUKLAB team's submission for SemEval-2025 Task 4: Unlearning Sensitive Content from Large Language Models. This task aims to selectively erase sensitive knowledge from large language models, avoiding both over-forgetting and under-forgetting issues. We propose an unlearning system that leverages Model Merging (specifically TIES-Merging), combining two specialized models into a more balanced unlearned model. Our system achieves competitive results, ranking second among 26 teams, with an online score of 0.944 for Task Aggregate and 0.487 for overall Aggregate. In this paper, we also conduct local experiments and perform a comprehensive analysis of the unlearning process, examining performance trajectories, loss dynamics, and weight perspectives, along with several supplementary experiments, to understand the effectiveness of our method. Furthermore, we analyze the shortcomings of our method and evaluation metrics, emphasizing that MIA scores and ROUGE-based metrics alone are insufficient to fully evaluate successful unlearning. Finally, we emphasize the need for more comprehensive evaluation methodologies and rethinking of unlearning objectives in future research. Code is available at https://github.com/zjunlp/unlearn/tree/main/semeval25.

Problem

• Selectively erase sensitive knowledge from large language models
• Avoid over-forgetting and under-forgetting issues in unlearning
• Evaluate unlearning effectiveness beyond MIA and ROUGE metrics

Innovation

• Uses Model Merging for unlearning sensitive content
• Applies TIES-Merging to balance model performance
• Analyzes unlearning via trajectories and weight dynamics