Verify robustness under matched context windows or burst-tolerant coding

Verify whether the robustness gap between WeaveMark and competing watermarking methods at high synonym-substitution rates is attributable to differences in context-window size or error-correction structure by matching context-window sizes across methods or replacing WeaveMark’s codes with a burst-tolerant alternative.

Background

The paper reports that WeaveMark is generally more robust than the compared methods under synonym-substitution attacks, but that Qu et al.’s method performs better at a substitution ratio of 0.3 for short messages. The authors hypothesize that this difference is partly caused by WeaveMark’s and BiMark’s larger context window, which causes a substitution to disrupt extraction at more subsequent positions, and partly by the symbol-level structure of Reed–Solomon coding used by Qu et al., which may better tolerate clustered bit errors.

The proposed explanations are not directly tested in the reported experiments. The unresolved task is therefore to isolate these factors through controlled comparisons using matched context-window sizes and alternative codes designed to tolerate burst errors.

References

Direct verification—for instance, by matching window sizes across methods or replacing our codes with a burst-tolerant alternative—is left for future work.

— WeaveMark: Robust and Scalable Multi-bit LLM Watermarking via Coded Payload Spreading  (2609.02177 - Park et al., 2 Sep 2026) in Appendix, Section “Additional robustness results,” subsection “Various message lengths and substitution ratios”