Sequential deletion requests degrade models because each one undoes the last, and a low-rank orthogonal projection stopped that
AspireAI community · arXiv (cs.LG)
Machine unlearning is usually evaluated as one request: delete this data, check the model forgot it and still works. Real deployments get those requests continuously, and this paper shows most unlearning methods fall apart when they arrive in sequence, because each new deletion can undo earlier ones and erode general utility at the same time. The proposed fix, Continual Orthogonal Projection, restricts each gradient update to the orthogonal complement of activation subspaces cached from prior forget sets, which the authors prove gives first-order non-interference with earlier deletions. It stores only a low-rank basis per layer, needs neither past gradients nor the previous forget sets, and plugs into existing gradient-based unlearning methods without changing their objective. On the TOFU and RWKU benchmarks with Llama-3.2-3B-Instruct it preserves more utility than the base objectives without weakening forgetting, and outperforms continual-unlearning baselines; the authors note that neither O3 nor ALKN maintains both forgetting and utility as requests accumulate. The full set of runs took about 300 GPU hours on one 96 GB workstation GPU, and the paper reports no headline effect size in its abstract.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.