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 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 paper introduces K-Bench, a benchmark designed to test whether unlearning holds up when language models act as agents rather than just answering questions directly. The authors argue that existing benchmarks like TOFU and MUSE certify forgetting only from a model's final response, so a model that simply declines to answer is treated as having forgotten the target knowledge. They show this model-level certification does not carry over to agentic deployments, where the model's behavior unfolds over multiple steps.
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.