Research map / Continual learning / Continual fine-tuning of LLMs
Machine unlearning by selective forgetting
Removes specific data influence from trained networks without full retraining.
24 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Continual fine-tuning of LLMs, keyword cluster “unlearning · multimodal · instruction”
State History Tried Checklist Preprints Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
6 papers in 2023–24 and 17 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 11.5% to 8.7%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2019 · Overcoming Multi-model Forgetting (ICML 2019 · 10 citations)
- 2023 · Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks Using an Incompetent Teacher (AAAI 2023 · 113 citations)
- 2024 · Fast Machine Unlearning without Retraining through Selective Synaptic Dampening (AAAI 2024 · 75 citations)
- 2025 · Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · 7 citations)
- 2026 · Remaining-data-free Machine Unlearning by Suppressing Sample Contribution (ICLR 2026 · 1 citation)
Earliest papers here
- Overcoming Multi-model Forgetting (ICML 2019 · 10 citations)
- Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks Using an Incompetent Teacher (AAAI 2023 · 113 citations)
- Boundary Unlearning: Rapid Forgetting of Deep Networks via Shifting the Decision Boundary (CVPR 2023 · 80 citations)
What has been tried
11 papers here report 109 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 11%, hurt 13%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Increase α from 10 to 20 | CIFAR100 rocket class unlearning · MIA | 2.20±0.00 -> 0.20 | Helped | Fast Machine Unlearning without Retraining through Selective Synaptic Dampening (AAAI 2024 · 75 citations) |
| Set α above the default to relax the forgetting condition | CIFAR-10, single-class forgetting · Accft | 0 -> 37.30; 0 -> 76.80; 0 -> 95.00 | Hurt | Decoupled Distillation to Erase: A General Unlearning Method for Any Class-centric Tasks (CVPR 2025 · 7 citations) |
| add paraphrased forget data | TOFU-10% · forget quality | 1e-7 -> 0.51 | Helped | Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · 4 citations) |
| Add classification loss preservation to feature distribution optimization | Age · NoMUS | 0.7021 -> 0.7536 | Helped | Distribution-Level Feature Distancing for Machine Unlearning: Towards a Better Trade-off Between Model Utility and Forgetting (AAAI 2025 · 2 citations) |
| add mask learning with penalized cross-entropy loss | CIFAR-100 · HM | 99.01 -> 99.05 | No clear effect | ESC: Erasing Space Concept for Knowledge Deletion (CVPR 2025 · 1 citation) |
| Add Inhibited Synthesis to PostFilter | SVHN · retaining accuracy (A_r) | 91.15±0.85 -> 92.68±0.13 | Helped | Toward Efficient Data-Free Unlearning (AAAI 2025) |
What reviewers will expect
Datasets and baselines named in the 5 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 3 of 5 papers
- CIFAR-100 · 3 of 5 papers
- Tiny ImageNet · 2 of 5 papers
Baselines
- SCRUB · 3 of 5 papers
Recent preprints
Loading recent arXiv preprints…
Use it
Download BibTeX of 24 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Continual pretraining of language models (66 papers)
- Mixture-of-experts for multimodal continual tuning (52 papers)
Elsewhere in the topic
- Low-rank and model-merging adaptation (90 papers)
- Synthetic data and merging for CLIP (32 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).