"Unified Parameter-Efficient Unlearning for LLMs"

Ever wondered how to make your LLM forget specific things without starting from scratch? Here's your answer.

LLMEraser introduces a unified framework for efficient unlearning in LLMs, enabling selective removal or correction of specific information while preserving overall model performance through parameter-efficient fine-tuning.

-----

https://arxiv.org/abs/2412.00383

🤔 Original Problem:

LLMs can inadvertently retain sensitive or incorrect information during fine-tuning. Current unlearning methods either require expensive retraining or struggle with precise information removal while maintaining model performance.

-----

🔧 Solution in this Paper:

→ LLMEraser introduces three categories of instance-wise unlearning: Instance Removal, Query Modification, and Response Correction.

→ It leverages influence functions to directly calculate parameter changes in PEFT adapters without retraining.

→ The framework reformulates parameter changes as a finite-sum quadratic programming problem for efficient computation.

→ It updates only the necessary adapter parameters while preserving the original model structure.

-----

💡 Key Insights:

→ Instance-wise unlearning can target specific data points without affecting related concepts

→ Parameter-efficient methods can achieve unlearning without full model retraining

→ Influence functions can effectively estimate parameter changes for selective forgetting

-----

📊 Results:

→ Performance gap of only 0.6% compared to full retraining

→ 5.1% average improvement over corrupted baselines in query modification tasks

→ Successfully handles 40% label noise in response correction tasks