Research map / Large language models / Language, safety and interpretability
WaterDrum: Watermark-based Data-centric Unlearning Metric
Xinyang Lu, Xinyuan Niu, Gregory Kang Ruey Lau, Nhung Bui, Rachael Hwee Ling Sim, John Russell Himawan, Fanyu Wen, Chuan-Sheng Foo, See-Kiong Ng, Bryan Kian Hsiang Low
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 watermark-based WaterDrum instead of ROUGE under exact duplicates in retain set | unlearning metric on retrained Llama-2-7B · WaterDrum-TOFU, exact duplicate | AUROC: 0.510 -> 0.926 | Improved |
| use WaterDrum instead of KnowMem | unlearning metric · WaterDrum-Ax, no duplicate | AUROC: 0.491 -> 0.965 | Improved |
| use WaterDrum instead of baseline metrics for calibration | unlearning metric · WaterDrum-Ax, exact duplicate | calibration R2: -37.47 (ROUGE) -> 0.987 | Improved |
| add semantic duplicates of the forget set to the retain set (ablation) | ROUGE metric · WaterDrum-TOFU | AUROC: 0.908 -> 0.798 | Hurt |
| add exact duplicates of the forget set to the retain set (ablation) | Truth Ratio metric · WaterDrum-TOFU | AUROC: 0.747 -> 0.508 | Hurt |
| add exact duplicates of the forget set to the retain set (ablation) | WaterDrum · WaterDrum-Ax | AUROC: 0.965 -> 0.957 | No clear effect |
| replace Waterfall with adapted KGW watermarking (ablation) | WaterDrum · WaterDrum-Ax | AUROC / calibration R2: AUROC 0.965 -> 0.871; R2 0.963 -> 0.996 | Mixed |
| replace Waterfall with adapted Synth-ID (ablation) | WaterDrum · WaterDrum-Ax | AUROC: 0.965 -> 0.549 | Hurt |
| replace Waterfall with adapted EXP-edit (ablation) | WaterDrum · WaterDrum-Ax | AUROC: 0.965 -> 0.789 | Hurt |
Closest papers by meaning
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.89)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.89)
- OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics (NeurIPS 2025 · similarity 0.88)
- MUSE: Machine Unlearning Six-Way Evaluation for Language Models (ICLR 2025 · similarity 0.88)
- Large Language Model Unlearning (NeurIPS 2024 · similarity 0.88)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.88)
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.
