Research map / Large language models / Language, safety and interpretability
LORD-GoF: A Robust Online Detection Approach for LLM Watermarks in Sparse and Mixed Streams
Jiade Xu, Zhouping Li
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 |
|---|---|---|---|
| replace fixed threshold alpha=0.05 with LORD online levels | GoF watermark detection · OPT-1.3B Gumbel-Max stream M=10000, pi=0.05 | FDR: about 0.50 -> near 0.05 target | Improved |
| replace fixed threshold with LORD in extremely sparse stream | GoF watermark detection · Qwen-2.5-3B Inverse Transform, pi=0.01 | FDR: 0.85 -> 0.00 | Improved |
| increase global sparsity pi toward 0.99 (dense stream) (ablation) | LORD-GoF · Qwen-2.5-3B Inverse Transform | power: converges to 1.0 | Improved |
| use Anderson-Darling statistic at tau=0.5 (ablation) | LORD-GoF · Sheared-LLaMA-2.7B Gumbel-Max | power / FDR: Power 0.854, FDR 0.000 | Improved |
| use Tr-GoF (Phi) instead of other GoF tests at tau=0.3 (ablation) | LORD-GoF · Sheared-LLaMA-2.7B Gumbel-Max | power / FDR: Power 0.839, FDR 0.000 | Improved |
| use sensitive GoF tests at tau=0.7 (ablation) | LORD-GoF · Sheared-LLaMA-2.7B Gumbel-Max | FDR: e.g. LORD-Chi FDR 0.094, LORD-Cra 0.079 | Hurt |
| use Phi test (ablation) | LORD-GoF · OPT-1.3B Gumbel-Max stream | power: Power not above 0.92 (And and Chi above 0.92) | Hurt |
| replace LORD with LORD++ (asymmetric reward) (ablation) | LORD-GoF · appendix comparison | power: absolute power gap at most 0.006 | No clear effect |
Closest papers by meaning
- On the Empirical Power of Goodness-of-Fit Tests in Watermark Detection (NeurIPS 2025 · similarity 0.92)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.88)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.88)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.87)
- On the Reliability of Watermarks for Large Language Models (ICLR 2024 · similarity 0.87)
- Provable Robust Watermarking for AI-Generated Text (ICLR 2024 · 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.
