The paper investigates how linear concept erasure methods fail when features are densely superposed, showing that such methods destroy both target and non‑target features. It demonstrates that networks trained with gradient descent can avoid this failure but converge to one of two distinct circuit‑level solutions—mirror or shadow—depending on initialization. The authors map this bifurcation as a function of feature entanglement, reveal a stable attractor structure, and use causal interventions to show that both solutions leave a measurable trace of the erased feature that can be recovered with a single scalar patch.
arXiv:2507. 07754v3 Announce Type: replace-cross Abstract: Machine unlearning is usually evaluated by what the classifier outputs: forget-set accuracy, confidence, membership-inference scores.
The paper introduces Unmerge, an efficient machine unlearning algorithm that treats unlearning as the inverse of task arithmetic. By representing the forget component as a low‑rank basis at each layer, Unmerge optimizes three goals—matching the merged vector, suppressing leakage, and bounding correction size—to limit forget leakage and retain damage. Experiments on ResNet‑50, ViT‑S/16, and Llama‑3.2‑3B show significant performance gains over existing methods while maintaining privacy and feature‑distribution fidelity.
arXiv:2601.22028v2 Announce Type: replace Abstract: Most LLM unlearning methods aim to approximate retrain-from-scratch behaviors with minimal distribution shift, often via alignment-style objectives...
arXiv:2608. 02058v1 Announce Type: new Abstract: Source-free class unlearning erases whole classes using only the forget data, judged at the representation level, where features can leak a class the head no longer predicts.
Approximate machine unlearning seeks to remove the influence of a forget set from a trained model without full retraining. Existing gradient-based methods require data-dependent hyperparameter search,...
The paper investigates whether representational entanglement—shared structure between knowledge domains—impedes unlearning in neural networks. By training six 254M‑parameter language models with varying degrees of disentanglement between biology and non‑biology knowledge and applying three unlearning methods, the authors find that more disentangled models consistently achieve better retain‑forget trade‑offs, with up to four‑fold lower retain cost. This controlled experiment provides direct evidence that entanglement contributes to collateral damage during unlearning, supporting a long‑standing hypothesis in interpretability research.
The paper investigates whether representational entanglement—shared structure between knowledge domains—impedes unlearning in neural networks. Using Selective Gradient Masking, the authors train six 254M‑parameter language models with varying degrees of disentanglement between biology and non‑biology knowledge, then apply three standard unlearning methods to each. Results show that more disentangled models consistently achieve better retain‑forget trade‑offs, with up to four‑fold lower retain cost at the same forgetting level, providing direct evidence that entanglement contributes to collateral damage in unlearning.
The paper investigates continual machine unlearning, where models must forget data over time. It identifies a fundamental issue called plasticity collapse, where successive unlearning requests cause geometric constraints that saturate parameter space, leading to two failure modes: forward failure (reduced forgetting quality) and backward failure (re‑memorization). Experiments across architectures and datasets confirm that plasticity collapse is a pervasive problem in continual unlearning.
The paper investigates how the structural entanglement of facts within a large language model’s knowledge base influences whether those facts leak after unlearning. Using two unlearning algorithms (WHP and GA+KL) across fictional and real-world datasets and multiple model sizes, the authors find that highly entangled facts are more likely to be recalled before unlearning, but the relationship changes—WHP weakens it while GA+KL reverses it. By directly manipulating entanglement scores and observing corresponding recall changes, they demonstrate a causal link and develop a predictive tool to audit prompts for potential leakage.
arXiv:2606. 20431v1 Announce Type: new Abstract: Continual learning (CL) systems often forget previously acquired knowledge, yet the mechanisms driving forgetting remain hard to isolate in practice because real datasets entangle many factors.
arXiv:2508.20443v3 Announce Type: replace Abstract: Large language models (LLMs) are trained on massive datasets that may include private or copyrighted content. Due to growing privacy and ownership...
arXiv:2609.06862v1 Announce Type: new Abstract: Superposition refers to neural networks representing more features than they have dimensions. It offers a possible explanation for polysemantic neurons...