Brief

Researchers have developed LineageMark, a novel white-box watermarking framework designed to trace contributions within multi-stage derivation chains of large language models (LLMs). This system encodes watermarks directly into model parameters using a projection-based approach with stable carriers to minimize sensitivity to model alterations. LineageMark is capable of preserving contributor watermarks through multiple derivation stages and supports incremental insertion by multiple users, demonstrating robustness against common modifications like fine-tuning, quantization, and pruning. AI