Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution

arXiv·high signal

Standard unlearning evaluations measure behavioral suppression at full precision immediately after training, but real deployments quantize models — where supposedly deleted knowledge can resurface. This paper introduces circuit attribution to identify and surgically ablate knowledge-encoding subnetworks, producing unlearning that persists through GPTQ, AWQ, and GGUF quantization pipelines. Critical for compliance with data deletion mandates (GDPR right-to-be-forgotten) in production-quantized LLM deployments.