Binomial Multibit LLM Watermark Achieves Superior Accuracy
A novel method for multibit LLM watermarking utilizing binomial encoding encodes each bit of the payload at all token positions. This technique, detailed in arXiv:2605.11653, features a stateful encoder that adjusts encoding pressure in real-time, focusing on bits that are underencoded during the generation process. When tested against eight baselines with payloads of up to 64 bits, this approach demonstrates enhanced message accuracy and resilience, particularly in scenarios involving large payloads and low distortion. Additionally, the study presents per-bit confidence scoring and critiques existing evaluation metrics for their insufficient practical relevance.
Key facts
- Proposes binomial encoding for multibit LLM watermarking.
- Encodes every bit of the payload at every token position.
- Includes a stateful encoder that redirects encoding pressure dynamically.
- Evaluated against 8 baselines on up to 64-bit payloads.
- Achieves superior message accuracy and robustness.
- Gap to baselines widens in large payload and low-distortion regimes.
- Introduces per-bit confidence scoring.
- Challenges prior evaluation metrics for lacking practical insights.
Entities
Institutions
- arXiv