Machine unlearning aims to remove the influence of private or copyrighted data from trained LLMs, a step toward regulatory compliance. But models are usually quantized before deployment to save memory and compute, and this compression can interfere with the unlearning process.
The arXiv preprint (2609.27355) studies this problem by looking at the loss landscapes for retained and forgotten data. The authors argue that the interaction between these landscapes determines whether unlearning survives quantization, and they propose a method that makes the forget direction more robust to weight perturbations introduced by quantization.
Since only the abstract is available, the exact mechanism and experimental results are not detailed here. The work points to a practical gap: unlearning that works on a full-precision model may not hold up in the compressed models that actually ship.