TL;DR
Training a model well past overfitting until it "grokks"—experiencing delayed generalization—drastically improves its ability to selectively unlearn data. Because grokking restructures internal representations into modular, disentangled subcircuits with nearly orthogonal gradient directions, post-grokking checkpoints enable faster unlearning, lower collateral damage to retained data, and greater stability without altering existing unlearning algorithms.
What Problem Does This Paper Address?
Machine unlearning seeks to erase the influence of specific training examples (e.g., for GDPR "right to be forgotten" compliance or data sanitation) without paying the enormous computational price of retraining from scratch. However, approximate unlearning methods often cause severe collateral damage, inadvertently degrading the model’s performance on retained data.
This failure stems from representational entanglement: when forget and retain data share the same feature representations and optimization pathways, deleting one inevitably damages the other. This raises a fundamental paradox around grokking (where models generalize long after memorizing the training set):
- Does extensive training cement memorized data so deeply that it becomes impossible to unlearn?
- Or does the cleaner, more structured representation formed after grokking make selective unlearning easier?
Methods
Rather than proposing a new unlearning algorithm, the authors systematically evaluate whether when you stop training (early-stopped $\theta_{\text{pre}}$ vs. post-grokking $\theta_{\text{grok}}$) affects downstream unlearning:
- Evaluated Unlearning Algorithms: Standard gradient ascent (GA), $\nabla\tau$ (interleaved ascent and descent), Fisher forgetting (curvature-guided noise), SCRUB (distillation-based unlearning), retain fine-tuning, and LLM-specific preference/representation algorithms (PO, NPO, RMU, KL).
- Controlled Baselines & Ablations:
- Compared grokked models against Sharpness-Aware Minimization (SAM) to isolate whether benefits come solely from flat loss basins.
- Compared against Max-Affine Spline models to isolate structural effects from baseline accuracy.
- Mechanistic Probes: Gradients between forget and retain sets (cosine similarity and angular separation), Centered Kernel Alignment (CKA) across layers, and Local Complexity (LC) via linear partition density.
- Local Grokking in LLMs: Because large language models rarely exhibit uniform, dataset-wide grokking, the authors introduce a metric to identify "locally grokked" examples (subsets of points that generalize early and remain stable) to compare their unlearning dynamics against ungrokked examples within the same model checkpoint.
Datasets
- Vision: CIFAR-10, SVHN, and ImageNet-100 (evaluated using ResNet-18 and standard 3-layer CNN architectures). Tested on both class-restricted subsets and random cross-class forget splits.
- Language: The TOFU (Task of Fictitious Unlearning) benchmark consisting of synthetic author profiles and question–answer pairs, fine-tuning the 1.3B-parameter Phi-1.5 model.
Key Insights and Ideas
- Representational benefits outweigh memorization depth: Despite being trained for hundreds of thousands of steps, grokked models unlearn significantly better and cleaner than early-stopped models.
- Orthogonal gradient pathways: In pre-grokked models, the gradient correlation (cosine similarity) between forget and retain sets is near-perfect ($\sim 0.990\text{–}0.999$, gradient angle $\approx 2.6^\circ$). In grokked models, gradient correlation plummets to $\sim 0.426\text{–}0.521$ (angle $\approx 64.8^\circ$). Updating parameters along the forget gradient barely interferes with retained knowledge.
- Modularity over simple flatness: CKA between retain and forget features drops by 72% (0.459 to 0.129) post-grokking, and local complexity drops threefold. While flat minima (via SAM) buffer against catastrophic collapse, SAM alone fails to match grokking, proving that architectural modularity—not just loss landscape flatness—is critical.
- Local grokking manifests in LLMs: Within a single Transformer, locally grokked facts exhibit 30–50% better unlearning selectivity and extraction attack resistance without hurting retain extraction strength.
Key Results
- CIFAR-10 (ResNet-18): Grokked models using SCRUB yielded an 8–9% increase in test accuracy alongside a 6–8% lower unlearning accuracy (UA) compared to pre-grokked models. Fisher Forgetting on grokked ResNets reached 99.42% retain accuracy.
- Convergence Speed: At a 15% forget rate, grokked models achieved target forgetting levels within 5–8 optimization steps compared to 15–20 steps for pre-grokked models (a 60–70% speedup).
- Robustness & Privacy: Grokked ResNets preserved superior post-unlearning adversarial robustness (18.1% vs. 11.2% accuracy under PGD attacks at $\epsilon=0.20$) and achieved Membership Inference Attack (MIA) scores closer to random guessing (e.g., 0.677 vs. 0.842 for SCRUB).
- TOFU (Phi-1.5): Locally grokked examples lowered unlearning Extraction Strength ($ES_{\text{unlearn}}$) under GA from 0.518 to 0.344 (a 34% improvement) while preserving retain score ($ES_{\text{retain}} = 0.605$ vs. $0.556$).
Limitations and Future Work
- Training Cost: Grokking requires 5–6$\times$ longer training, making it computationally impractical to train full models to grokking purely for unlearning convenience.
- Scope & Scale: Vision setups relied on relatively small architectures (CNN, ResNet-18), and language experiments were limited to Phi-1.5 on a synthetic dataset.
- Heuristic Identification: "Local grokking" currently relies on heuristic loss thresholding between checkpoints rather than a formal mechanistic criterion.
- Future Work: Identifying lightweight optimization techniques or regularization schemes that directly induce modularity and gradient orthogonality early in training, delivering grokking-level unlearning benefits without the massive compute overhead.
Summary updated