Research map / Large language models / Instruction and fine-tuning
Rethinking Benign Relearning: Syntax as the Hidden Driver of Unlearning Failures
Sangyeon Yoon, Hyesoo Hong, Wonje Jeung, Albert No
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 |
|---|---|---|---|
| unlearn with syntactically diversified (paraphrased) forget set instead of original | GA/NPO unlearning on TOFU forget05 · TOFU Real Authors | utility average: 0.4014 -> 0.4852 | Improved |
| unlearn with syntactically diversified forget set | unlearning on TOFU forget05 · TOFU Retain set | utility average: 0.1607 -> 0.3128 | Improved |
| unlearn with syntactically diversified forget set | unlearning on TOFU forget05 · TOFU World Facts | utility average: 0.6056 -> 0.6104 | Improved |
| unlearn with syntactically diversified forget set | unlearning on TOFU forget05 · TOFU World Facts | Probability: 0.4187 -> 0.4169 | No clear effect |
| use diversified forget set to reduce syntactic similarity to the syntactic relearn set | unlearning on TOFU · TOFU forget05 | syntactic similarity: 0.4513 -> 0.2241 | Improved |
| unlearn with diversified forget set | unlearning on TOFU · TOFU forget05 relearning with syntactic set | relearn success rate | Improved |
| use syntactically similar relearn set instead of topically relevant set | GA unlearned model · TOFU forget05 | relearn success rate | Hurt |
| use SCRUB instead of GA/NPO for unlearning | unlearning on TOFU · TOFU forget05 | relearn success rate | Hurt |
| evaluate relearning at every step with equal step budget (best step) instead of one epoch | BLUR relearning evaluation · WMDP/WHP/RWKU | ROUGE-L | No clear effect |
| relearn with LoRA instead of full parameters on fully unlearned model | TOFU unlearned model · TOFU forget05 | relearn success rate | Hurt |
| use safety training (DPO) instead of unlearning algorithms | unlearning · TOFU syntactic relearning | relearn vulnerability | Hurt |
| relearn on perfectly retrained model with topic or syntactic set | retrain model · TOFU | relearn success rate: 0% -> 0% | No clear effect |
Closest papers by meaning
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.89)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.89)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.89)
- RA$^2$: Retain-Anchored Attraction Preserves Forget-Adjacent Utility in LLM Unlearning (NeurIPS 2026 · similarity 0.88)
- What to Forget in Unlearning? Forget Set Curation for Language Models (NeurIPS 2026 · similarity 0.88)
- RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 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.
