Research map / Large language models / Language, safety and interpretability
Towards Effective Evaluations and Comparisons for LLM Unlearning Methods
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.
Proposes Unlearning with Control (UWC), an evaluation framework that selects extraction strength as a robust unlearning metric and calibrates retention via model mixing so unlearning methods can be fairly compared on TOFU.
Problem
Current LLM unlearning metrics are susceptible to red-teaming attacks and do not reflect true knowledge removal, and comparing methods is confounded by the trade-off between removal and retention.
Approach
Use extraction strength (ES) as the basic metric, chosen for robustness via Pearson correlation under jail-breaking, probing, relearning and token noising; calibrate retention post-unlearning by mixing pre- and post-unlearning parameters (model mixing) to equalize non-target ES, then compare ES on targeted data.
Key results
- After UWC calibration on TOFU 1% setup, NPO achieves the lowest ES-exact unlearn score on Phi-1.5 (0.1259) while GD gives the lowest on Llama-2-7B (0.0293), indicating GA-style methods remain competitive with tuning.
- ES was identified as the most robust metric, showing the highest Pearson correlation under relearning and embedding probing attacks versus PPL, EM, and ROUGE.
- Temperature scaling (TS) improved KL unlearning, e.g., Llama-2-7B 5% ES-exact unlearn dropping from 0.0953 to 0.0730 in the 10% setting, and TS yields the lowest ES-perturb unlearn score of 0.0000 in one setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use extraction strength (ES) instead of ROUGE, PPL, or EM as the unlearning metric | LLM unlearning evaluation · Red-teaming attacks: relearning and embedding probing | Pearson correlation coefficient | Helped |
| Calibrate retention with model mixing (MM) before comparing removal | LLM unlearning method comparisons · TOFU | ES on targeted data | Helped |
| Use the 5% rather than 1% or 10% unlearning setup | TOFU unlearning · TOFU | unlearning task difficulty | Hurt |
| Add a retention loss to gradient ascent (GA), using GD or KL regularization | GA-based unlearning · TOFU | ES on non-targeted data | Helped |
| Tune the loss-selection proportion (LS) for KL (ablation) | KL · TOFU | ES-exact unlearn: 0.2123 -> 0.3540; 0.4912 -> 0.4347 | Mixed |
| Add batch-size adjustment (BS) to KL (ablation) | KL · TOFU | ES-exact unlearn: 0.4912 -> 0.3358; 0.2791 -> 0.2287 | Mixed |
| Add early stopping (ES) to KL (ablation) | KL · TOFU | ES-exact unlearn: 0.2123 -> 0.2033; 0.4912 -> 0.2163 | Mixed |
| Add temperature scaling (TS) to KL (ablation) | KL · TOFU | ES-exact unlearn: 0.4912 -> 0.0730 | Helped |
| Adjust the learning rate (LR) for KL (ablation) | KL · TOFU | ES-exact unlearn: 0.2123 -> 0.2031; 0.4912 -> 0.4585 | Mixed |
Surprising findings
- Larger forget rates do not necessarily yield harder unlearning; the 5% setup was more challenging than 1% and 10%.
- Without calibration, GA and RMU collapse to near-zero ES on both retain and unlearn sets, showing excessive unlearning that destroys retention.
Limitations
- Metric robustness is assessed on only two LLMs and the TOFU benchmark.
- Calibration requires the non-targeted data and a tuned threshold tau; results depend on tau choice.
- Gold-standard retrained models are not used, so the true extent of removal cannot be validated directly.
Experimental setup
Datasets: TOFU
Benchmarks: TOFU
Metrics: extraction strength (ES-exact) ES-perturb perplexity ROUGE-L exact memorization KL divergence Pearson correlation coefficient
Baselines: GA GD KL PO NPO RMU
Builds on: gradient difference negative preference optimization representation misdirection for unlearning tuned lens TOFU
Closest papers by meaning
- OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics (NeurIPS 2025 · similarity 0.95)
- Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness (NeurIPS 2025 · similarity 0.94)
- Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond (ICLR 2025 · similarity 0.94)
- WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models (NeurIPS 2024 · similarity 0.93)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.93)
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.93)
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.