Suhang Wang

This paper introduces 'unlearning corruption attacks,' demonstrating that the performance degradation inherent in approximate graph unlearning can be exploited by an adversary to significantly reduce the accuracy of Graph Neural Networks (GNNs) after targeted data deletion.

The paper argues that current LLM benchmark datasets are often contaminated by being included in pretraining data, and proposes that future benchmarks must be contamination-resistant and support inference to maintain reliable model evaluation.

The paper distinguishes between a model's ability to generate useful updates for external agent components (harness-updating) and its ability to benefit from those updates (harness-benefit), finding that updating capabilities are surprisingly uniform while benefit is maximized in mid-tier models.