Research map / Large language models / Language, safety and interpretability
Selective Forgetting: Advancing Machine Unlearning Techniques and Evaluation in Language Models
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 SeUL, a selective span-level machine unlearning method for LMs that penalizes only sensitive spans instead of fully reversing the loss, plus S-EL/S-MA metrics and online/offline sensitive span annotation.
Problem
Machine unlearning of sensitive/personal information in language models, where existing fully reversed-loss unlearning (KUL) degrades generation ability and fails under adversarial injection of common knowledge.
Approach
Instead of negating the loss over all tokens of a forget instance, SeUL applies the reversed likelihood only to selected sensitive spans, chosen either online by low language-model probability or offline via a two-stage LLM (ChatGPT) annotation with backward verification; it also introduces S-EL and S-MA metrics that only count sensitive tokens.
Key results
- SeUL on GPT-Neo 2.7B reaches S-EL10 of 0.3 vs 1.7 for KUL and S-MA of 14.6 vs 23.0 at matched EL/MA thresholds, with classification accuracy 50.2 vs 49.9.
- On dialogue tasks, SeUL achieves average PPL 179.8 vs 539.1 for KUL on GPT-Neo 125M (F1 7.8 vs 1.1).
- On Mistral-7B, SeUL gets average accuracy 55.7 vs 53.3 for KUL and S-MA 20.9 vs 23.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use selective span-level unlearning instead of KUL's full loss negation | GPT-Neo 125M · Forget set (d=32) | S-EL₁₀ and S-MA: S-EL$_{10}$: 1.3 -> 0.3; S-MA: 22.9 -> 16.1 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | GPT-Neo 1.3B · Forget set (d=32) | S-EL₁₀ and S-MA: S-EL$_{10}$: 1.4 -> 1.9; S-MA: 23.9 -> 23.6 | Mixed |
| Use selective span-level unlearning instead of KUL's full loss negation | GPT-Neo 2.7B · Forget set (d=32) | S-EL₁₀ and S-MA: S-EL$_{10}$: 1.7 -> 1.0; S-MA: 23.0 -> 14.6 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | Llama2 7B · Forget set (d=32) | S-EL₁₀ and S-MA: S-EL$_{10}$: 1.5 -> 1.2; S-MA: 22.3 -> 12.3 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | Mistral 7B · Forget set (d=32) | S-EL₁₀ and S-MA: S-EL$_{10}$: 2.1 -> 1.3; S-MA: 23.6 -> 20.9 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | GPT-Neo 125M · Avg. 4 dialogue datasets | F1: 1.1 -> 7.8 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | GPT-Neo 1.3B · Avg. 4 dialogue datasets | F1: 10.4 -> 11.4 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | GPT-Neo 2.7B · Avg. 4 dialogue datasets | F1: 10.1 -> 10.9 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | Llama2 7B · Avg. 4 dialogue datasets | F1: 9.7 -> 10.4 | Helped |
| Use selective span-level unlearning instead of KUL's full loss negation | Mistral 7B · Avg. 4 dialogue datasets | F1: 13.3 -> 12.8 | Hurt |
| Use offline annotated sensitive spans instead of online-selected spans (ablation) | SeUL on GPT-Neo 125M · Avg. 8 classification datasets | accuracy: 41.0 -> 41.5 | Helped |
| Use offline annotated sensitive spans instead of online-selected spans (ablation) | SeUL on GPT-Neo 125M · Avg. 4 dialogue datasets | PPL: 179.8 -> 102.4 | Helped |
Surprising findings
- Under the injected common-knowledge attack, KUL cannot continue a commonsense sentence after unlearning while SeUL can.
Limitations
- Online selection introduces inaccurate spans compared with offline annotation.
- Evaluation of sensitive spans relies on ChatGPT-based annotation.
- Smaller models show less stable performance under unlearning.
Experimental setup
Datasets: Training Data Extraction Challenge (Pile) forget set Hellaswag Winogrande COPA ARC-Easy ARC-Challenge PIQA MathQA PubMedQA Wizard of Wikipedia Empathetic Dialogues Blended Skill Talk
Benchmarks: Hellaswag Winogrande COPA ARC-Easy ARC-Challenge PIQA MathQA PubMedQA Wizard of Wikipedia Empathetic Dialogues Blended Skill Talk Wizard of Internet
Metrics: S-EL S-MA EL MA accuracy F1 perplexity
Baselines: KUL NER-based annotation
Builds on: KUL (knowledge unlearning) extraction likelihood memorization accuracy
Closest papers by meaning
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.90)
- What to Forget in Unlearning? Forget Set Curation for Language Models (NeurIPS 2026 · similarity 0.90)
- Mitigating Memorization in Language Models (ICLR 2025 · similarity 0.90)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.90)
- Constrained Entropic Unlearning: A Primal-Dual Framework for Large Language Models (NeurIPS 2025 · similarity 0.90)
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.