Research map / Large language models / Language, safety and interpretability

What to Forget in Unlearning? Forget Set Curation for Language Models

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 0 citations (OpenAlex)

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.

Studies forget set curation for LLM unlearning, introducing the CleanSlate benchmark for verbatim output suppression over songs and books, and finds that retrieval-based curators give weak selective suppression while an evaluation-aware curator forgets nearly everything but causes collateral damage.

Problem

Unlearning methods assume the forget set is already known, but in practice a suppression request for a song or book must first be mapped to the corpus data used for unlearning; this upstream curation step is unstudied.

Approach

Frame forget set curation as a separate pipeline stage (request to forget set to unlearnt model), measure corpus footprints of works via n-gram coverage, and evaluate curators end-to-end with per-model extraction profiles, content-grounded QA and capability retention.

Key results

• Evaluation-aware curator with SimNPO reaches 99.4% average forget-side reduction but 86.3% retain-side reduction and an average QA drop of -6.5 pp.
• BM25-pre curator with SimNPO yields only 21.4% average forget reduction and 23.4% retain reduction, showing weak suppression.
• With BM25-pre forget sets, average forget reduction is 10.5% under SimNPO, 98.2% under UNDIAL, and 32.8% under RMU, showing the unlearner strongly changes outcomes.

Surprising findings

• Strong forgetting under any tested curator-unlearner configuration was always accompanied by substantial retain-side suppression; no configuration was selective.
• Under the EA curator, UNDIAL keeps average validation accuracy roughly unchanged (+0.1 pp) while SimNPO drops it by about 5 pp on identical forget inputs.

Limitations

• Narrow setting: verbatim output suppression for songs and books only; does not cover paraphrase, translation, or semantic reuse.
• Exact-overlap diagnostics miss paraphrase and semantic references; corpus footprint measures are lower bounds and do not establish causal responsibility.
• Retrieval curator and EA results evaluated at fixed operating points and fixed hyperparameters; no configuration was found to be selective.

Experimental setup

Datasets: Dolmino mix (dolma3_dolmino_mix-10B-1025) Dolma3 pretraining subset Common Crawl Jan 2025 Billboard Hot 100 songs matched to LRCLib Project Gutenberg books GSM8K BBH WinoGrande CoQA HumanEval+ LAMBADA HellaSwag

Benchmarks: CleanSlate CleanSlate-QA GSM8K BBH WinoGrande CoQA HumanEval+ LAMBADA

Metrics: net reduction in extractable windows (forget and retain) QA accuracy change (pp) validation accuracy change (pp)

Baselines: BM25 pre BM25 mid Infini-gram mid evaluation-aware curator

Builds on: SimNPO RMU UNDIAL Infini-gram probabilistic extraction

Closest papers by meaning

• Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.92)
• SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion (NeurIPS 2026 · similarity 0.91)
• A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.91)
• Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.91)
• LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.91)
• RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 2025 · similarity 0.91)