TokenSwap is a novel approach that selectively replaces token probabilities to prevent LLMs from reproducing copyrighted content without affecting overall performance.

• Works as a post-hoc solution without requiring model retraining
• Targets grammar-related tokens to disrupt memorized sequences
• Preserves model performance while reducing verbatim content generation
• Addresses legal and security concerns without extensive computational resources

This research addresses critical security challenges in AI deployment by protecting intellectual property and reducing legal exposure without compromising model utility.

A Lightweight Method to Disrupt Memorized Sequences in LLM