Tag
The study reveals substantial gaps in machine unlearning for LLMs, showing that adversarial evaluation uncovers recoverability of forgotten information despite strong standard metrics, highlighting the need for adversarial stress-testing.
This paper identifies tool-mediated recovery as a failure mode in LLM unlearning and proposes Agentic Tool Unlearning (ATU) to reduce both parametric recall and tool-based recovery while preserving normal tool use.
The paper introduces ConceptGuard, a benchmark for evaluating context-sensitive unlearning in large language models using dual-use concepts, revealing that current unlearning techniques perform poorly under this practical evaluation framework.
This paper explores forget set curation for machine unlearning in language models, introducing a benchmark to evaluate verbatim output suppression and highlighting trade-offs between effectiveness and capability retention.
The paper proposes AdaPop, an adaptive popularity-based method for LLM unlearning that adjusts gradient pressure based on fact frequency to improve forgetting effectiveness and reduce leakage under queries.
This paper identifies a privacy vulnerability in RL-trained multimodal large reasoning models, which can leak sensitive facts in their reasoning traces even after unlearning, and proposes LEMUR, a training-free inference-time framework that uses entropy dynamics to detect and suppress such leakage.
Proposes J-Access, an inference-time audit using the Jacobian lens to measure residual knowledge accessibility in unlearned LLMs, finding that accessibility predicts recovery speed but that directly minimizing it fails to promote genuine deletion.
Introduces GROM, a gradient-free one-shot machine unlearning method that computes a closed-form additive weight update via ridge-regularized least squares, achieving state-of-the-art forgetting-utility trade-offs on benchmarks like TOFU and WMDP, and resisting quantization-based recovery attacks.
This paper introduces a new benchmark for evaluating machine unlearning, focusing on multi-hop reasoning consistency and recovery robustness. Experiments show existing unlearning methods face an 'impossible triangle' trade-off among forget quality, robustness, and utility preservation.
This paper proposes a retain-aware localization method for machine unlearning that reduces collateral damage to semantically similar retained examples, and introduces a retain-similar evaluation set. Experiments on CIFAR-10 with ResNet18 show reduced collateral damage and improved unlearning metrics.
This paper investigates exact deletion from language-model memory, showing that subtractive methods work when record influence is addressable, while replay/rebuild is needed when influence is woven into recurrent state. Experiments on Gemma and Kimi hybrid models demonstrate trade-offs in utility and exactness.
Introduces LENS, a contextualization-based evaluation protocol for testing narrative unlearning in large language models, evaluating suppression across direct, attributed, contrastive, and abstract levels.
This paper introduces Stochastic Meta-Unlearning (SMU), a bilevel framework that uses VLM-level feedback to learn an unlearning-ready initialization for the language backbone, achieving better forget-retain trade-offs in multimodal unlearning.
OriginBlame is a record- and token-level data provenance system that propagates author identity through AI training data pipelines, enabling precise forget sets for machine unlearning. It eliminates over-deletion from dataset-level systems and improves unlearning effectiveness.
Proposes GSUO, a guidance-signal-aware optimization framework for machine unlearning that uses fine-grained signals to guide the forgetting process, avoiding over-unlearning and under-unlearning, and outperforms 14 baselines.
This paper frames machine unlearning in LLMs as an asymmetric generalization problem, introduces the SUITE evaluation protocol and training corpus to address under- and over-forgetting, and presents JensUn++, an algorithm achieving the best forget-retain utility trade-off across three LLMs.
A comprehensive survey of methods, datasets, and benchmarks for multimodal unlearning across vision, language, video, and audio, providing a taxonomy and highlighting open problems.
Proposes a practical auditor that uses membership inference attacks to compute data-dependent lower bounds on the unlearning parameter, finding a sharp separation between certified algorithms (e.g., model clipping, rewind-to-delete) that achieve tight bounds and empirical methods (e.g., Hessian-based unlearning, gradient ascent) that exhibit large bounds, indicating poor unlearning.
CBD introduces an API-only black-box unlearning framework for LLMs that uses two auxiliary models to create controlled behavioral divergence between retained and target data, achieving a better unlearning-utility trade-off compared to existing methods.
This position paper argues that the term 'machine unlearning' is overused in LLM research, advocating for stricter terminology tied to dataset-defined deletion and retraining-equivalence guarantees.