Optimal Multi-bit Generative Watermarking Characterization via Linear Programming
November 2, 2026
New encoding-decoding constructions achieve the theoretical lower bound for multi-bit watermarking in LLMs under worst-case false-alarm constraints. The approach formulates watermark design as a linear program, identifying failure mechanisms in previous suboptimal schemes to characterize optimal performance.
HOW THIS AFFECTS YOU
●
researcherYou can use these linear programming formulations to design watermarking schemes that meet theoretical optimality bounds.
●
policyThis provides a more rigorous mathematical foundation for verifying the provenance of generative AI content.