<ul data-eligibleForWebStory="true"><li>Large language models trained on web data may include problematic datapoints.</li><li>Complete retraining to eliminate such datapoints is computationally prohibitive.</li><li>A new algorithm called MSA proposes an efficient way to estimate and undo the influence of individual datapoints by leveraging model checkpoints.</li><li>Experimental results show that MSA outperforms existing machine unlearning algorithms, suggesting it could lead to more flexible large language models capable of data erasure.</li></ul>

Model State Arithmetic for Machine Unlearning

Discover more