A new arXiv preprint introduces GRIN+, a machine unlearning approach aimed at removing patient data from trained medical models both quickly and effectively. The work focuses on imbalanced clinical datasets, a setting where existing unlearning methods tend to degrade model performance. It frames the problem around privacy rules such as GDPR and HIPAA that grant patients the right to have their data erased.
A new arXiv paper examines the gap between a model appearing to forget a deleted record and its recurrent state still retaining that record's influence. The authors store the state difference produced right after a record is written and track it through subsequent processing steps, deriving a condition under which the omission is exact along with its computational cost and a replay-based certificate. The work targets attention mechanisms with recurrent memory, where deleting data from context does not necessarily remove its effect from internal state.
A new arXiv preprint examines how large audio-language models can be made to forget sensitive information they may have memorized during training. The work focuses on the speech question-answering setting, where such models have shown strong performance but also carry privacy risks. The authors frame machine unlearning as a way to reduce unintended retention of private data in these systems.
A new arXiv preprint addresses the computational cost of data attribution, which measures how individual training samples affect a model's behavior. The authors frame the problem around estimating the influence matrix at scale, with applications in data valuation, machine unlearning, and interpretability. The abstract highlights that scaling such methods has remained a longstanding obstacle.
This paper examines how to verifiably delete stored items from a memory that places a support-vector boundary around keys and weights their values by the resulting coefficients. It shows that when a coefficient is exactly zero, the key can be removed without altering the current normal, and identifies the conditions under which such deletion guarantees stop holding. The work aims toward auditable removal in memory-based learning systems.
The paper installs a support-vector gate inside a frozen Gemma 3 model and records which stored keys and values correspond to each exchange, allowing a deletion request to be traced to specific memory entries. It argues that an assistant can stop repeating a fact without actually removing it, so the work separates genuine deletion from mere output suppression. The authors present this as a step toward deletion requests that can be audited rather than taken on trust.
A new audit examines 263 publicly released checkpoints that use batch normalization and finds that reported unlearning results shift depending on which version of those statistics is used. Because batch-norm statistics are not produced by gradient updates and are rarely documented in model releases, refitting them on retained data can change the numbers an evaluation relies on. The authors argue that this makes some unlearning verdicts unreliable, since apparent forgetting may reflect checkpoint bookkeeping rather than the removed data genuinely being gone.
A new arXiv paper proposes an intrusion detection approach for digitised electrical distribution networks that combines neuromorphic temporal embeddings with a hybrid spiking neural network and XGBoost classifier. The authors frame the work as a response to the high computational cost of existing deep-learning-based detection systems. They also evaluate robustness when machine unlearning attacks are used against the model.
Researchers have proposed a machine unlearning technique that directs the removal of memorized information to specific layers of a large language model rather than treating the whole network. Framed as a lighter-weight alternative to full retraining, the approach seeks to erase sensitive, copyrighted, or otherwise undesirable training content while preserving the model's remaining capabilities.