CoLU preprint reports lower memory use for selective AI unlearning

Researchers report a method for making language models forget selected information that uses 77.4% to 79.5% less GPU memory than full-model unlearning across measured settings. Their CoLU preprint describes updating one selected model layer while balancing removal of targeted information against preservation of other knowledge. A screening procedure ranks layers through brief trials on a subset of the forgetting data, then examines the strongest candidates using the full data. The authors report that screening reduces CoLU runtime by 60.2-79.7% with almost the same performance. CoLU improved both forgetting and retention over three comparison methods in all six TOFU benchmark settings, and over SimNPO in five of six. The findings are benchmark results reported in a preprint.