Research map / Large language models / Language, safety and interpretability
Selective Disclosure Watermarking for Large Language Models
Xuyang Chen, Xiang Li, Yangxinyu Xie, Qi Long
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use single-level HeRo instead of two-level (drop selective disclosure) at same payload (ablation) | HeRo · C4, 24-bit payload | bit accuracy: 0.925 -> 0.981 | Improved |
| hierarchical vocabulary routing instead of BiMark (20) for multi-bit watermark | HeRo1 (3) · C4, 24-bit payload | bit accuracy: 0.894 -> 0.981 | Improved |
| two-level HeRo (selective disclosure) compared with MPAC (2, 2.0) | HeRo2 (2,2) · C4, 24-bit payload | bit accuracy: 0.882 -> 0.925 | Improved |
| two-level HeRo (1,1) compared with MPAC (1, 2.0) | HeRo2 (1,1) · C4, 48-bit payload | bit accuracy: 0.800 -> 0.798 | No clear effect |
| use HeRo instead of MPAC to preserve text quality | HeRo1 (3) · C4, 24-bit payload | median perplexity: 4.76 -> 4.33 | Improved |
| unbiased hierarchical sampling vs unwatermarked text | HeRo2 (2,2) · C4, 24-bit payload | median perplexity: 4.31 -> 4.36 | No clear effect |
| increase number of first-layer chunks K1 (public level) (ablation) | HeRo2 · C4, 24-bit payload | public vs private bit accuracy | Mixed |
| public-only verifier decoding the private payload (ablation) | HeRo2 · C4, 24-bit payload | private bit accuracy: approximately 50% | No clear effect |
| batched hierarchical implementation instead of MPAC/BiMark released code for decoding | HeRo1 (3) · decoding latency, batch size 1 | decoding latency (us/token): 130.43 -> 7.75 | Improved |
| increase batch size for HeRo decoding (ablation) | HeRo1 (3) · decoding latency | decoding latency (us/token): 7.75 -> 1.44 | Improved |
| increase batch size for BiMark decoding (ablation) | BiMark (20) · decoding latency | decoding latency (us/token): 130.43 -> 133.23 | No clear effect |
| increase random replacement rate p (ablation) | HeRo1 (3) · C4, 24-bit payload | bit accuracy: 0.949 -> 0.752 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- BiMark (20) — BiMark: Unbiased Multilayer Watermarking for Large Language Models (ICML 2025) · better on 6 of 7 shared results
- StealthInk (1) — StealthInk: A Multi-bit and Stealthy Watermark for Large Language Models (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- BiMark: Unbiased Multilayer Watermarking for Large Language Models (ICML 2025 · similarity 0.88)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.88)
- StealthInk: A Multi-bit and Stealthy Watermark for Large Language Models (ICML 2025 · similarity 0.87)
- Majority Bit-Aware Watermarking for Large Language Models (NeurIPS 2026 · similarity 0.87)
- Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization (NeurIPS 2025 · similarity 0.87)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
