Enterprises fine-tune language models on proprietary data that may later require removal due to privacy, contractual, or compliance obligations. Selective unlearning removes requested knowledge while preserving model utility, offering a practical alternative to full retraining, but existing methods treat the explicitly identified forget examples as the complete deletion scope. This is insufficient when target knowledge remains recoverable through paraphrases, aliases, or neighboring training examples. We propose GRAPHSU, a graph-guided controller that expands the deletion scope beyond forget seeds by constructing a weighted support-route graph, propagating deletion pressure through it, and applying graded forgetting strengths to high-risk neighbors. On the Task of Fictitious Unlearning (TOFU), a synthetic author-profile question-answering benchmark, and PISTOL, a structural-unlearning benchmark built around interconnected factual samples, with GPT-2 Medium and Llama-3.2-3B-Instruct, GRAPHSU achieves the lowest utility-feasible soft leakage across all deletion settings, reducing leakage by up to 49.5 percentage points over a matched seed-only baseline, demonstrating that effective enterprise unlearning requires controlling support routes, not just forget seeds.
Waqas Khan, Tabinda Sarwar, Jingyue Cong et al.· 0 citations
Evolutionary algorithms are well suited to neural architecture search and other combinatorial design problems, but their scalability is often limited by the high cost of fitness evaluation. This paper studies evolutionary multi-objective optimisation in a regime where fitness evaluations are effectively free, enabled by a training-free proxy for neural network expressivity. We propose SWAP-Lite, a Pareto-guided evolutionary algorithm that maintains an explicit archive of non-dominated solutions over representational capacity and deployment cost, yielding an anytime optimiser that exposes budget-feasible solutions throughout the search. Using a MobileNet-style architecture space as a case study, we instantiate the fitness function with a sample-wise activation pattern proxy and perform large-scale evolutionary searches with up to 105 architecture evaluations. Experiments on CIFAR-10 and ImageNet show that SWAP-Lite discovers compact architectures that are competitive with state-of-the-art training-based and zero-shot baselines, while reducing search cost by one to four orders of magnitude. Analysis of the evolutionary dynamics demonstrates that explicit bi-objective optimisation produces higher-quality constrained Pareto fronts and superior anytime hypervolume compared with random search, greedy local search, and single-objective evolutionary baselines.
Jingyue Cong, Kai Wei, Yameng Peng et al.· Annual Conference on Genetic...· 0 citations