Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution

TL;DR

This paper introduces MANSU, a novel unlearning method that ensures forgetting persists after quantization by combining circuit attribution, null-space projection, and quantization-aware bounds, addressing systematic failures in existing approaches.

Contribution

The paper proposes MANSU, a new unlearning technique that guarantees persistent forgetting post-quantization, and introduces Circuit Attribution Divergence as a verification metric for structural erasure.

Findings

MANSU outperforms gradient-based methods under quantization.

Existing methods fail to maintain forgetting after 4-bit quantization.

MANSU achieves meaningful forgetting and structural erasure across models.

Abstract

Standard unlearning evaluations measure behavioral suppression in full precision, immediately after training, despite every deployed language model being quantized first. Recent work has shown that 4-bit post-training quantization can reverse machine unlearning; we show this is not a tuning artefact but a systematic dual failure: gradient-based methods that achieve meaningful forgetting lose it under compression, while methods that survive quantization barely change the model. Both failures trace to the same root cause: across all baselines, per-parameter updates lie 47-828x below the NF4 quantization bin width; updates diffused across billions of parameters cannot clear quantization bin boundaries, a consequence we formalize as a sparsity-permanence tradeoff. We present MANSU (Mechanistic-Aligned Null-Space Unlearning), which resolves both modes by combining causal circuit attribution to isolate the minimal forget-set subgraph, circuit-restricted null-space projection with a diagonal-Fisher retain bound, and a per-parameter magnitude floor guaranteeing quantization survival by construction. We additionally introduce Circuit Attribution Divergence (CAD), a mechanistic verification metric distinguishing structural erasure from behavioral suppression, a distinction existing metrics cannot make. Across multiple model families and hazard benchmarks, MANSU is the first method to jointly satisfy all four properties with margin on each (meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure), while gradient-based baselines recover up to +0.05 accuracy under compression.