Research map / Large language models / Language, safety and interpretability
OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces OpenUnlearning, a unified extensible framework for benchmarking LLM unlearning methods and metrics, plus a meta-evaluation of unlearning metrics for faithfulness and robustness and a benchmark of 8 methods on TOFU.
Problem
LLM unlearning research is fragmented across benchmarks, methods, and evaluation metrics, and the reliability of unlearning metrics is unclear, hindering comparison and reproducibility.
Approach
A modular, config-driven library that integrates 13 unlearning algorithms, 16 metrics, and 3 benchmarks (TOFU, MUSE, WMDP), together with a meta-evaluation that scores metrics on faithfulness (AUC over known-positive/negative model pools) and robustness (quantization and relearning stress tests).
Key results
- Extraction Strength is the most reliable metric in meta-evaluation, with overall aggregate 0.85 (faithfulness 0.92, robustness 0.79).
- On TOFU benchmarking, SimNPO ranks first with aggregate 0.53, preserving utility at 1.00 and privacy at 0.63.
- Truth Ratio has the highest faithfulness (0.95) but ranks lower overall due to weaker robustness (0.59).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SimNPO instead of the other evaluated unlearning methods | TOFU unlearning · TOFU | Agg.: 0.53 (SimNPO); 0.52 (RMU) | Helped |
| Use SimNPO to balance memorization, privacy, and utility | TOFU unlearning · TOFU | Utility: 1.00 | Helped |
| Use RMU rather than SimNPO | TOFU unlearning · TOFU | Utility: 0.61 (RMU); 1.00 (SimNPO) | Hurt |
| Over-unlearn with GradDiff | TOFU unlearning · TOFU | Priv.: 3e-3 | Hurt |
| Under-unlearn with IdkNLL | TOFU unlearning · TOFU | Priv.: 0.17 | Hurt |
| Use Extraction Strength rather than the other evaluated metrics | TOFU metric evaluation · TOFU metric meta-evaluation | Agg.: 0.85 | Helped |
| Use Truth Ratio rather than less faithful metrics | TOFU metric evaluation · TOFU metric meta-evaluation | Faithful.: 0.95 | Helped |
| Use MIA-based metrics as the sole unlearning assessment | TOFU metric evaluation · TOFU metric meta-evaluation | Robustness Agg.: 0.57 (MIA - ZLib); 0.52 (MIA - MinK); 0.52 (MIA - LOSS); 0.48 (MIA - MinK++) | Hurt |
| Apply 4-bit quantization when evaluating ROUGE | ROUGE metric · TOFU metric meta-evaluation | Quantization robustness: 0.93 | No clear effect |
| Relearn the forget-set data when evaluating ROUGE | ROUGE metric · TOFU metric meta-evaluation | Relearn robustness: 0.48 | Hurt |
Surprising findings
- MIA-based privacy metrics show high faithfulness but low robustness, so they are unreliable on their own for assessing unlearning.
- Different ranking schemes produce very different method rankings, indicating sensitivity of benchmark conclusions to ranking choices.
Limitations
- Findings rest on the TOFU setup, which makes simplifying assumptions about data distribution and target models.
- Relies on popular benchmarks that have been called weak measures of unlearning progress.
- Chosen desiderata (faithfulness and robustness) may not be a comprehensive set for good unlearning metrics.
Experimental setup
Datasets: TOFU MUSE WMDP
Benchmarks: TOFU MUSE WMDP
Metrics: Extraction Strength Exact Memorization Truth Ratio Paraphrased Probability MIA (ZLib, MinK, LOSS, MinK++) Model Utility Forget Fluency Faithfulness (AUC-ROC) Robustness (relearning, quantization) harmonic mean aggregate
Baselines: GradDiff IdkDPO IdkNLL NPO SimNPO RMU UNDIAL AltPO
Builds on: TOFU MUSE WMDP NPO SimNPO
Closest papers by meaning
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.95)
- Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.93)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.93)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.92)
- WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models (NeurIPS 2024 · similarity 0.92)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.92)
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.