Research map / Large language models / Language, safety and interpretability
Learning-Time Encoding Shapes Unlearning in LLMs
Ruihan Wu, Konstantin Garov, Kamalika Chaudhuri
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 |
|---|---|---|---|
| fine-tune with paraphrased descriptions (FT-Mul) instead of single descriptions | unlearning with GA and TV · Eval-DU+ and TOFU+, extraction trade-off | Norm-AUC / AUC: FT-Mul outperforms or matches FT-Single in 44 out of 48 cases | Improved |
| fine-tune with paraphrased descriptions (FT-Mul) | TV UL-Exact, Llama2-7B · Eval-DU+ extraction | Norm-AUC: 0.57 -> 0.65 | Improved |
| fine-tune with paraphrased descriptions (FT-Mul) | GA UL-Exact, Llama2-7B · Eval-DU+ extraction | Norm-AUC: 0.59 -> 0.55 | Hurt |
| fine-tune with paraphrased descriptions (FT-Mul) | TV UL-Mul, Llama2-7B · Eval-DU+ extraction | Norm-AUC: 0.69 -> 0.72 | Improved |
| fine-tune with paraphrased descriptions (FT-Mul) | GA UL-Exact, Llama3-8B · Eval-DU+ extraction | Norm-AUC: 0.55 -> 0.62 | Improved |
| fine-tune with paraphrased descriptions (FT-Mul) | GA UL-Exact, Gemma2-2B · Eval-DU+ extraction | Norm-AUC: 0.52 -> 0.60 | Improved |
| fine-tune with paraphrased descriptions (FT-Mul) | FT-Mul versus FT-Single, memorization trade-off · Eval-DU+ and TOFU+, memorization trade-off | Norm-AUC / AUC: outperforms or matches FT-Single in 39 out of 48 cases | Improved |
| align the unlearn split with chunk boundaries of the training text | FT-Mul-Chunk, GA UL-Exact, Llama2-7B · Eval-DU+ extraction | Norm-AUC: 0.53 -> 0.66 (+0.13) | Improved |
| align the unlearn split with chunk boundaries | FT-Mul-Chunk, TV UL-Exact, Gemma2-2B · Eval-DU+ extraction | Norm-AUC: 0.61 -> 0.63 (+0.02) | Improved |
| isolate target knowledge by concatenating individual descriptions into chunks (FT-Mul-Chunk-Iso) | GA UL-Single, Llama2-7B · Eval-DU+ extraction | Norm-AUC: +0.11 | Improved |
| isolate target knowledge in chunks (FT-Mul-Chunk-Iso) | TV UL-Mul, Llama2-7B · Eval-DU+ extraction | Norm-AUC: +0.13 | Improved |
| organize unlearn targets by author partitions (TOFU+) rather than entangled chunks (Eval-DU+) | GA UL-Exact, Llama2-7B · extraction trade-off | Norm-AUC: 0.53 -> 0.59 | Improved |
Closest papers by meaning
- Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness (NeurIPS 2025 · similarity 0.93)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.92)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.92)
- Probing Hidden Knowledge Holes in Unlearned LLMs (NeurIPS 2025 · similarity 0.92)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.91)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.91)
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.
