Learning-Time Encoding Shapes Unlearning in LLMs

2025/06/18 by Ruihan Wu, Wu, Ruihan, Konstantin Garov +3
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Natural Language Processing Techniques

paper · pdf · doi:10.48550/arxiv.2506.15076

openalex publication_date 2025/06/18 · openalex created_date 2025/10/19 · openalex updated_date 2026/07/28

Abstract

As large language models (LLMs) are increasingly deployed in the real world, the ability to ``unlearn'', or remove specific pieces of knowledge post hoc, has become essential for a variety of reasons ranging from privacy regulations to correcting outdated or harmful content. Prior work has proposed unlearning benchmarks and algorithms, and has typically assumed that the training process and the target model are fixed. In this work, we empirically investigate how learning-time choices in knowledge encoding impact the effectiveness of unlearning factual knowledge. Our experiments reveal two key findings: (1) learning with paraphrased descriptions improves unlearning performance and (2) unlearning individual piece of knowledge from a chunk of text is challenging. Our results suggest that learning-time knowledge encoding may play a central role in enabling reliable post-hoc unlearning.

Citations

• Deep Contrastive Unlearning for Language Models
• PEBench: A Fictitious Dataset to Benchmark Machine Unlearning for Multimodal Large Language Models
• UIPE: Enhancing LLM Unlearning by Removing Knowledge Related to Forgetting Targets
• Beyond Single-Value Metrics: Evaluating and Enhancing LLM Unlearning with Cognitive Diagnosis
• WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models
• Evaluating Deep Unlearning in Large Language Models
• Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization
• Position: LLM Unlearning Benchmarks are Weak Measures of Progress
• Towards Robust Knowledge Unlearning: An Adversarial Framework for Assessing and Improving Unlearning Robustness in Large Language Models
• The Llama 3 Herd of Models
• Gemma 2: Improving Open Language Models at a Practical Size
• Revisiting Who's Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective
• MUSE: Machine Unlearning Six-Way Evaluation for Language Models
• To Forget or Not? Towards Practical Knowledge Unlearning for Large Language Models
• Avoiding Copyright Infringement via Large Language Model Unlearning
• RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models
• Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference
• Large Language Model Unlearning via Embedding-Corrupted Prompts
• Large Scale Knowledge Washing
• To Each (Textual Sequence) Its Own: Improving Memorized-Data Unlearning in Large Language Models
• SOUL: Unlocking the Power of Second-Order Optimization for LLM Unlearning
• Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
• Guardrail Baselines for Unlearning in LLMs
• The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning
• Eight Methods to Evaluate Robust Unlearning in LLMs
• Machine Unlearning of Pre-trained Large Language Models
• Rethinking Machine Unlearning for Large Language Models
• TOFU: A Task of Fictitious Unlearning for LLMs
• Knowledge Unlearning for LLMs: Tasks, Methods, and Challenges
• DEPN: Detecting and Editing Privacy Neurons in Pretrained Language Models
• Large Language Model Unlearning
• In-Context Unlearning: Language Models as Few Shot Unlearners
• Who's Harry Potter? Approximate Unlearning in LLMs
• Can Sensitive Information Be Deleted From LLMs? Objectives for Defending Against Extraction Attacks
• Physics of Language Models: Part 3.2, Knowledge Manipulation
• Physics of Language Models: Part 3.1, Knowledge Storage and Extraction
• Knowledge Sanitization of Large Language Models
• Llama 2: Open Foundation and Fine-Tuned Chat Models
• Composing Parameter-Efficient Modules with Arithmetic Operations
• Direct Preference Optimization: Your Language Model is Secretly a Reward Model
• GPT-4 Technical Report
• On Provable Copyright Protection for Generative Models
• Editing Models with Task Arithmetic
• Knowledge Unlearning for Mitigating Privacy Risks in Language Models
• Continual Learning and Private Unlearning
• Training language models to follow instructions with human feedback
• Locating and Editing Factual Associations in GPT
• Measuring Massive Multitask Language Understanding
• Machine Unlearning
• PIQA: Reasoning about Physical Commonsense in Natural Language
• RACE: Large-scale ReAding Comprehension Dataset From Examinations
• Atyaephyra at SemEval-2025 Task 4: Low-Rank Negative Preference Optimization
• Opt-Out: Investigating Entity-Level Unlearning for Large Language Models via Optimal Transport

Related

• Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness
• Unlearning That Lasts: Utility-Preserving, Robust, and Almost Irreversible Forgetting in LLMs
• UCD: Unlearning in LLMs via Contrastive Decoding
• Understanding the Dilemma of Unlearning for Large Language Models
• UIPE: Enhancing LLM Unlearning by Removing Knowledge Related to Forgetting Targets