Machine unlearning aims to remove specific knowledge from a pre-trained model—such as sensitive personal data or harmful behaviors—without retraining from scratch. However, recent research has revealed a disturbing trend: unlearning is often "skin deep." A model that appears to have forgotten a concept can often be forced to remember it through a few steps of fine-tuning or even simple model quantization.
The emerging consensus across recent work is that this fragility is a geometric problem. To make unlearning stick, we must move beyond simply changing the model's outputs and instead fundamentally reshape its loss landscape.
The Fragility of Sharp Minima
When we "unlearn" a data point using standard gradient descent, we are essentially looking for a new set of weights where the loss on the "forget" data is high. The problem, as identified by Chongyu Fan, Jinghan Jia, and colleagues in Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond, is that standard optimizers often land the model in sharp minima.
In a sharp minimum, the model's "forgotten" state exists in a narrow, precarious valley. Because the valley is so thin, a tiny perturbation in the model's parameters—caused by a few steps of new training or even a "jailbreak" prompt—can push the model back into a region where it remembers the original data. This is why models often suffer from "relapses" during downstream tasks, a phenomenon explored in diffusion models by Boheng Li and others in Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning.
Seeking Stability in Flat Basins
To solve this, researchers are turning to Sharpness-Aware Minimization (SAM) and related smoothness constraints. The goal is to steer the model toward flat minima: wide, stable basins where the loss remains consistently high (for forgotten data) or consistently low (for retained data) even if the parameters are slightly shifted.
Zheyuan Liu and Han Yan propose a framework called PRISM in Dual-Space Smoothness for Robust and Balanced LLM Unlearning. PRISM enforces "smoothness" in two places:
- Parameter Space: It ensures that small weight updates don't suddenly restore forgotten knowledge.
- Representation Space: It ensures that the model's internal activations don't contain "hidden" traces of the deleted data.
By finding these flat regions, the "relearn margin" is expanded. An attacker would need significantly more data and compute to "climb out" of a wide basin of forgetting than they would to escape a narrow, sharp one.
Why Noisy Optimization Helps
Counter-intuitively, making the optimization process worse can make the unlearning better. Yicheng Lang, Sijia Liu, and their team show in Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning that high-precision optimizers are actually a liability.
Standard high-precision gradients are very good at finding those tiny, sharp holes in the landscape. By "downgrading" to noisier, lower-precision optimizers (like zeroth-order or sign-based methods), the optimization process is physically unable to settle into sharp minima. The noise forces the model to find the largest, flattest regions of the landscape, which are inherently more robust to post-training modifications like quantization.
Targeted Geometry: Parameter Selection
We don't necessarily need to smooth the entire model. Li Xiong 0001 and Hong-Kyu Lee suggest a more surgical approach in Sharpness-Aware Parameter Selection for Machine Unlearning. By calculating the Hessian diagonal (a measure of curvature), they identify specific "salient" parameters that live in flat regions. Updating only these parameters ensures that the unlearning is both computationally efficient and naturally resistant to being undone.
The Dual Role of Sharpness
Interestingly, sharpness can be used as both a shield and a sword. In Sharpness-Aware Machine Unlearning, Haoran Tang and Rajiv Khanna introduce Sharp MinMax. This algorithm uses SAM to keep the "retain" data in a flat minimum (preserving generalization) while simultaneously using sharpness maximization to aggressively overfit and erase the "forget" data.
This highlights the core intuition of the field: Unlearning is a battle over the geometry of the weight space. By intentionally navigating toward flat basins, researchers are moving away from fragile, "temporary" forgetting toward a state of permanent, resilient erasure.