UNDO improves the compute–robustness Pareto frontier for unlearning on synthetic language and arithmetic tasks.

The claim comes from the 2025 preprint "Distillation Robustifies Unlearning" (arXiv 2506.06278). That paper introduces UNDO (Unlearn, Noise, Distill on Outputs). The method first applies an ordinary unlearning method to a model. It then perturbs a copy of that model's weights by mixing them with noise, and distills the unlearned model's outputs into the perturbed copy. More noise and more distillation compute make the result harder to relearn.

Read operationally, "improves the compute–robustness Pareto frontier" means this. Compute is the cost of the procedure, measured as a share of the original training compute. Robustness is how well the unlearned capability resists a relearning attack, meaning fine-tuning on the forgotten data. In this space, UNDO's points dominate those of the baselines. The baselines are unlearning methods used alone (gradient-difference and entropy-maximization style methods, plus robustness-oriented variants) and full retraining on filtered data, which is the most robust but costs the full training compute. UNDO fills in the middle of the curve: it pays part of the compute and gets part of the robustness that retraining gives.

The evidence is the authors' own experiments. They are small in scale, on synthetic language tasks and arithmetic tasks, and as far as is known they have not been independently replicated. The claim is limited to those tasks, so it does not depend on larger benchmarks. The paper reports more limited results on a realistic hazardous-knowledge benchmark. The frontier result depends on which baselines and relearning attacks were tested. A stronger attack, or a robustness-tuned baseline left out of the comparison, could change it. Within the stated scope, the reported results support the claim. How general it is beyond that scope remains open.