GROM Enables One-Shot Machine Unlearning via Closed-Form Analytical Updates
August 7, 2026
GROM replaces iterative, gradient-based unlearning with a ridge-regularized least-squares optimization. This approach provides a direct analytical solution for weight updates, preventing the 'hidden knowledge' issue where quantization can restore erased data.
HOW THIS AFFECTS YOU
●
researcherYou can achieve more robust unlearning without the high compute costs of fine-tuning.
●
policyThis offers a more reliable method for complying with data deletion and privacy regulations.