cs.AISep 25, 2026
SaveNeuralyzing the Trace: Selective Representation-Level Unlearning with Contrastive Sparse Autoencoders
Organizations: Mila · ILLS · École normale supérieure Paris-Saclay · IRT Saint Exupéry
Abstract
Machine unlearning aims to remove targeted information while preserving a model's other abilities. In realistic settings, such as privacy requests under the EU GDPR, the target may be narrow, for example information associated with a single person. Behavioral forgetting alone may be insufficient, motivating interventions directly on internal representations. However, standard mechanistic-interpretability extractors are poorly selective for such targets. We identify an energy bias in reconstruction-based extraction, which favors dominant background structure over low-energy target-specific components. We introduce SCALPEL, a contrastive sparse autoencoder designed to learn more selective forget features. We show theoretically that contrastive training promotes target-selective features and that our selection score controls expected background knowledge perturbation. We validate SCALPEL experimentally on TOFU across Qwen, Llama, and Gemma, where it substantially improves over NMF and standard SAE interventions and is competitive with Gradient Difference and RMU, bridging mechanistic interpretability and fine-grained unlearning.
Figures & tables
Figure 1: Overview of the SCALPEL pipeline. 1 We generate multiple semantic views of each forget anchor, then 2 cache residual-stream activations at layer . 3 An anchor-contrastive SAE learns sparse features using same-anchor views as positives and retain samples (purple) as negatives, so that forget samples (green) become isolated from retain information in the sparse feature space. 4 We select forget-specific features, decode their contribution , and subtract it from the residual stream before the remaining layers produce the prediction. The same selected feature set is applied to every input, while subtraction occurs only when those features activate. The intervention therefore acts as an input-dependent internal filter intended to suppress forget-related components while limiting its effect on retain activations.
Figure 2: Energy comparison between dominant background structure and background-orthogonal forget residuals at layer of Qwen. Left: background activations projected onto the first three principal components learned from retain and Wiki activations. Right: forget activations after removing these directions, shown in a second PCA basis fitted to the residuals. Since the panels use different bases, they should be compared through their scale and explained energy, not their coordinates. The background top-3 PCA energy is larger.
Figure 3: Unlearning trade-off on TOFU forget10 for Qwen2.5-1.5B-Instruct , after method-specific hyperparameter optimization and aggregation over five independent end-to-end seeds. Each curve varies the intervention strength: lower forget probability indicates stronger forgetting, while higher model utility indicates better preservation. Results for Llama and Gemma are reported in Appendix F.1 are similar. For more metrics look at Appendix F.2 .
Figure 4: Diagnostics linking representation geometry, feature selectivity, and intervention quality for SCALPEL . Left: contrastive training increases within-target alignment and reduces target–background overlap in the norm-aware sparse code. Right: features with higher selectivity scores generally achieve better single-feature forget–retain trade-offs.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: UMAP projection ( McInnes et al., 2018 ) of residual-stream activations from the forget, retain, and Wiki sets. Forget and retain examples strongly overlap because they come from the same generation process and differ only through the random assignment of authors to the splits: the full forget set is not a globally separable concept in activation space.
| Hyperparameter | Range / choices | Sampling | Applies to |
|---|---|---|---|
| Latent dimension | categorical | all | |
| Selection threshold | uniform | all | |
| Batch size | categorical | trained SAEs | |
| Training epochs | int. uniform | trained SAEs | |
| Encoder depth | categorical | trained SAEs | |
| Sparsity coefficient | log-uniform | trained SAEs |
Table 1: SAE-family search space (per trial). “Trained SAEs” means all methods except NMF; “contrastive family” means all contrastive variants.
| Hyperparameter | Range / choices | Sampling | Method |
|---|---|---|---|
| Learning rate | log-uniform | both | |
| Training epochs | categorical | both | |
| Forget-loss weight ( gamma ) | log-uniform | both | |
| Retain-loss weight ( alpha ) | log-uniform | both | |
| Retain loss type | categorical | GradDiff | |
| Steering coefficient | log-uniform | RMU |
Table 2: Trainer search space (GradDiff and RMU). The names gamma and alpha are the trainer implementation’s loss weights and are unrelated to the paper’s intervention strength and norm exponent .
| Method | Qwen2.5 | Llama3.2 | Gemma2 |
|---|---|---|---|
| Contrastive SAE | |||
| SCALPEL -Split | |||
| Standard SAE | |||
| NMF | |||
| GradDiff | |||
| RMU |
Table 3: Per-study outcomes: completed trials / pruned trials / best margin (Eq. equation 11 ).
Figure 6: Unlearning trade-off on TOFU forget10 for Qwen2.5-1.5B-Instruct , after method-specific hyperparameter optimization and aggregation over five independent end-to-end seeds. Each curve varies the intervention strength: lower forget probability indicates stronger forgetting, while higher model utility indicates better preservation. Results for Llama and Gemma are reported in Appendix F.1 .
Figure 7: Unlearning trade-off on TOFU forget10 for Llama-3.2-1B-Instruct , after method-specific hyperparameter optimization and aggregation over five independent end-to-end seeds. Each curve varies the intervention strength: lower forget probability indicates stronger forgetting, while higher model utility indicates better preservation.
Figure 8: Unlearning trade-off on TOFU forget10 for Gemma-2-2B-it , after method-specific hyperparameter optimization and aggregation over five independent end-to-end seeds. Each curve varies the intervention strength: lower forget probability indicates stronger forgetting, while higher model utility indicates better preservation.
Figure 9: Multi-metric forgetting–utility trade-offs on TOFU forget10 for Qwen2.5-1.5B-Instruct , after method-specific hyperparameter optimization. Curves show the mean over five finetuning seeds, and shaded regions show the seed-wise minimum–maximum range. Columns correspond to different forget-side metrics, while rows report preservation through model utility, retain Q–A probability, and retain Q–A ROUGE, for which higher values are better. Lower forget Q–A ROUGE and forget Q–A probability indicate stronger forgetting. Forget truth ratio and extraction strength are interpreted relative to the retain-only reference, while privacy leakage is best when close to zero. The black star and red cross denote the full and retain-only reference models, respectively; RMU and GradDiff are shown at their selected operating points.
Figure 10: Multi-metric forgetting–utility trade-offs on TOFU forget10 for Llama-3.2-1B-Instruct , after method-specific hyperparameter optimization. Curves show the mean over five finetuning seeds, and shaded regions show the seed-wise minimum–maximum range. Columns correspond to different forget-side metrics, while rows report preservation through model utility, retain Q–A probability, and retain Q–A ROUGE, for which higher values are better. Lower forget Q–A ROUGE and forget Q–A probability indicate stronger forgetting. Forget truth ratio and extraction strength are interpreted relative to the retain-only reference, while privacy leakage is best when close to zero. The black star and red cross denote the full and retain-only reference models, respectively; RMU and GradDiff are shown at their selected operating points.
Figure 11: Multi-metric forgetting–utility trade-offs on TOFU forget10 for Gemma-2-2B-it , after method-specific hyperparameter optimization. Curves show the mean over five finetuning seeds, and shaded regions show the seed-wise minimum–maximum range. Columns correspond to different forget-side metrics, while rows report preservation through model utility, retain Q–A probability, and retain Q–A ROUGE, for which higher values are better. Lower forget Q–A ROUGE and forget Q–A probability indicate stronger forgetting. Forget truth ratio and extraction strength are interpreted relative to the retain-only reference, while privacy leakage is best when close to zero. The black star and red cross denote the full and retain-only reference models, respectively; RMU and GradDiff are shown at their selected operating points.
Figure 12: Per-author top-5 feature selectivity for the reconstruction-only SAE, contrastive SAE, and SCALPEL -Split, evaluated across the five Qwen benchmark seeds. Points correspond to the mean top-5 selectivity score for each author, while the shaded region summarizes variation across authors. Adding contrastive learning increases feature selectivity over the standard SAE, and the split architecture further amplifies this effect.
More in this topic
Machine Unlearning
Machine unlearning removes the influence of selected data or concepts from trained models while preserving useful behavior and avoiding full retraining.
See all 351 papers