Proposing the Update-Matching Unlearning Method UMU
Start here
An arXiv paper released on October 15, 2026, proposes UMU, an update-matched unlearning method for scenarios where large language models continue to be fine-tuned after release, aiming to resist relearning of deleted knowledge while preserving model plasticity. The authors propose the reference timing hypothesis: when target knowledge suppression is comparable, a reference model that has undergone the same simulated relearning updates preserves plasticity better than a pre-update reference. UMU applies the same simulated relearning updates to both the defended model and a reference model that was never defended and initialized from the same unlearned model, using new-task learning to guide model selection. In experiments on TOFU and FaithUn datasets with Llama and Qwen models, UMU performed better than three baselines.
Generated by AI from reports · Updated 28 days ago
Attention and developments
Attention, hourly
Now 29, 0% in 24 hours·Peak 29 (Nov 6 03:00)
Related research
This event's paper
- Not All Change Is Damage: Update-Matched Unlearning for Large Language Models
arXiv · 2026-10-15 · Deep reading
Progress timeline
- arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesNot All Changes Are Harmful: Update-Matched Unlearning for Large Language Models
The paper proposes the update-matched unlearning method UMU for scenarios where models continue to be fine-tuned after release, preserving plasticity while resisting re-learning of deleted knowledge. The authors propose the reference timing hypothesis: when target knowledge suppression is comparable, a reference model that has undergone the same simulated re-learning updates preserves plasticity better than a pre-update reference. UMU applies the same simulated re-learning updates to the defended model and a reference model that was never defended and initialized from the same unlearned model, using new-task learning to guide model selection. In experiments on TOFU and FaithUn with Llama and Qwen models, UMU reduces harmful and benign re-learning by an average of 30.4 and 9.9 percentage points over three base unlearning methods, and improves new-task accuracy by 3.5 percentage points; when the reference model does not receive the updates, re-learning is similarly low, but new-task accuracy is lower than baseline methods, which the authors claim supports the reference timing hypothesis.