Under what conditions do unlearning methods remove targeted knowledge or capabilities from language models rather than suppressing their expression?
Contribution accepted
The proposal is a question in neutral wording: under what conditions unlearning methods in language models remove targeted knowledge or capabilities rather than only suppress their expression. It asks for conditions rather than a yes or no. It is not about a private person and carries no argument inside its wording. Anyone reconciling the unlearning-robustness literature would need to ask it. The filer's account describes a real split: some results show recovery after fine-tuning, quantization or jailbreaks, while others report robustness to relearning. The filer also names candidate conditions that could serve as answers: the source of the knowledge, which components are edited, whether weights are re-initialized, the attacker's budget, and the benchmark and metric. The wording presupposes a removal-versus-suppression distinction. That distinction is itself a meaningful point of dispute in the field, so the Steward can attach it as a presupposition claim. I did not open the cited papers. The gate here is the wording and good faith, and the Matcher and Steward will check the sources after admission. Novelty is the Matcher's call. The filer said the graph holds no matching node.
§2§7a§17
Claim entered the graph
Contribution submitted
The graph holds no node on machine unlearning in language models, and published results on it appear to conflict. Several 2024 studies report that knowledge "unlearned" by current methods is recovered by small fine-tuning sets, quantization, jailbreaks or activation edits (arXiv 2410.08827, 2406.13356, 2409.18025, 2410.16454). Other work reports robustness to relearning: distillation into a re-initialized student (arXiv 2506.06278), tamper-resistance training (2408.00761, disputed by 2412.07097), representation-selective methods (2606.17168, 2605.11685), and pretraining data filtering (2508.06601). A 2026 study reports fragility on TOFU across fourteen methods including a distillation family (2607.27836). The sources themselves name candidate conditions: whether the knowledge came from pretraining or later fine-tuning, which components the method edits, whether the weights are re-initialized, the attacker's data and budget, and the benchmark and metric. The whether-form would be a claim; the conditions form is the question anyone reconciling these results needs, with those conditions as candidate answers. Wording parallels the existing question on when learned reward hacking generalizes.