From Domains to Instances: Dual-Granularity Data Synthesis for LLM Unlearning

2026-01-07

arXiv.org

AI Summary

Existing benchmarks for unlearning in large language models (LLMs) struggle to distinguish between domain-level and instance-level forgetting and fail to accurately assess real-world unlearning efficacy. To address this, this work proposes BiForget, a framework that formally defines unlearning at two granularities for the first time. BiForget intrinsically leverages the target model itself—via seed-guided and adversarial prompting—to generate highly relevant and diverse unlearning data without relying on external generators. The method establishes an automated pipeline for synthetic data generation and evaluation, achieving approximately 20% higher relevance and 0.05 greater diversity compared to existing approaches on benchmarks such as Harry Potter, while reducing data volume by 50%. Furthermore, it yields more robust unlearning performance and better preserves model utility.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Generation
• Search and Optimization: Learning to Search

Application Category

• 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
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Although machine unlearning is essential for removing private, harmful, or copyrighted content from LLMs, current benchmarks often fail to faithfully represent the true"forgetting scope"learned by the model. We formalize two distinct unlearning granularities, domain-level and instance-level, and propose BiForget, an automated framework for synthesizing high-quality forget sets. Unlike prior work relying on external generators, BiForget exploits the target model per se to elicit data that matches its internal knowledge distribution through seed-guided and adversarial prompting. Our experiments across diverse benchmarks show that it achieves a superior balance of relevance, diversity, and efficiency. Quantitatively, in the Harry Potter domain, it improves relevance by ${\sim}20$ and diversity by ${\sim}$0.05 while halving the total data size compared to SOTAs. Ultimately, it facilitates more robust forgetting and better utility preservation, providing a more rigorous foundation for evaluating LLM unlearning.

Problem

• machine unlearning
• LLM unlearning
• forgetting scope
• data synthesis
• unlearning granularity

Innovation

• LLM unlearning
• dual-granularity
• data synthesis
• adversarial prompting
• forget set generation