Memorization-heavy models suffer greater damage during unlearning
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Researchers have uncovered a fundamental tension in model training that directly impacts unlearning effectiveness. Models that rely heavily on memorization suffer greater performance degradation when unlearning is applied to retain sets, compared to generalization-focused models. Using grokking experiments in modular addition as a controlled testbed, the work demonstrates that learning strategy shapes unlearning vulnerability. This finding matters for practitioners building systems requiring selective capability removal, from safety alignment to privacy compliance. The bucketed modular addition framework enables fine-grained control over memorization versus generalization contributions, offering a new diagnostic tool for understanding how training dynamics propagate through the unlearning pipeline.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper isolates a specific causal link: models trained via memorization don't just unlearn worse on target data, they also degrade more on retain sets. This asymmetry suggests unlearning isn't a simple erasure problem but a representation-level consequence of how knowledge was encoded during training.
This connects directly to the October 6th finding on spurious forgetting (reference 2), which showed that finetuning doesn't erase representations but shifts them. Here, the authors demonstrate that memorization-heavy models have less representational flexibility when unlearning is applied, making them more brittle. It also contextualizes the language-based unlearning vulnerabilities from late September (reference 1): if memorized facts are encoded more rigidly, they're harder to selectively remove without collateral damage. The grokking testbed offers the controlled setting that the multilingual and finetuning papers lacked, making this a mechanistic foundation for why unlearning fails in practice.
If practitioners applying this framework to real LLMs find that models trained on high-memorization corpora (e.g., raw web text) suffer greater retain-set degradation than those trained on curated, reasoning-heavy data, the finding generalizes beyond toy problems. If not, the result may be specific to modular arithmetic's structure.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
Language models hide rather than erase old knowledge during finetuning
Researchers have identified why language models appear to forget old knowledge during finetuning, yet often retain it in recoverable form. The work reveals that finetuning shifts representations along shared directions rather than erasing them, and that normalization mechanisms can restore access to buried facts once new training stabilizes. This mechanistic insight matters for practitioners: it…
Mentionsgrokking · modular addition · unlearning · memorization · generalization
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “How Learning Governs Unlearning across the Memorization-Generalization Spectrum”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.