SynMark watermarks LLM code via semantic coordinates, not token prefixes
Original title: SynMark: Semantic-Coordinate Multi-Bit Watermarking during LLM Code Generation
After full identifier renaming, BitAcc stays at 82.27% on CodeLlama and 83.77% on DeepSeek-Coder.
The paper proposes SynMark, a multi-bit watermark for LLM-generated code meant to survive identifier renaming and structural refactoring. Scope learns semantic coordinates from aligned prefixes of functionally equivalent variants and applies a syntax-aware adaptive bias at generation. Traced recomputes routes from transformed code and aggregates token contributions by reliability. Across three code language models, the authors report gains of up to 4.53 and 11.18 percentage points in clean BitAcc and Pass@1 over the strongest watermarked baselines.
Unconfirmed: Accuracy and Pass@1 figures are the authors' reported results, not independently verified.
- Surface-token routing can lose sync after functionality-preserving edits.
- Under 100% identifier renaming, CodeLlama BitAcc is 82.27%.
- DeepSeek-Coder BitAcc is 83.77% under the same renaming condition.
- Clean BitAcc and Pass@1 rise by up to 4.53 and 11.18 points.
From the original article
1 min read at arxiv.orgLLM-generated code is increasingly reused across downstream applications and often undergoes identifier renaming or structural refactoring, making robustness to functionality-preserving edits essential for generative multi-bit code watermarking. Existing methods commonly derive message routes from surface-token prefixes, so such edits can desynchronize routing and degrade watermark recovery.
By Rui Xu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.