- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies continual concept erasure, where new erasure requests arrive over time and the model must remove new targets without undoing prior erasures or degrading unrelated generation. The authors identify a cross-edit interference accumulation problem in existing closed-form editing methods: residual perturbations outside the retain set interact and accumulate across sequential updates, degrading general generation quality and sometimes collapsing previously erased targets into noise. They propose CEASE, a training-free method that imposes two subspace constraints on a closed-form solver: AIC adds the shared replacement token representation to the invariance matrix, and HOC projects the current update onto the orthogonal complement of dominant historical output directions when interference is detected. A theoretical decomposition attributes interference to repeated activation of the shared replacement and overlap between successive update directions. Experiments on continual erasure of celebrities, artistic styles, and instances show that CEASE achieves a more consistent erase–preserve trade-off.
Critical strengths
- The problem is clearly defined, the continual erasure setting is practically relevant, and the authors explicitly point out the limitation of existing methods that "protect only a finite retain set."
- The theoretical analysis is solid: based on the rank-one update structure, it derives a noise decomposition into anchor, off-anchor, and cross terms, and designs AIC and HOC accordingly, with tight motivation–design correspondence.
- The method is training-free and closed-form, with low computational overhead and easy integration into existing editing frameworks.
- Experiments cover three task types and multiple backbones, with comprehensive metrics (erase set, retain set, COCO), sufficient ablation and layer-scope analysis, and convincing qualitative results.
Critical weaknesses
- The theoretical decomposition relies on rank-one approximation and ideal constraint conditions (e.g., βᵀc*=0); in practice, numerical precision and regularization terms may break this structure, leaving a gap between theoretical guarantees and actual behavior.
- HOC uses the SVD of cumulative historical updates as a proxy for historical directions, but the paper admits the column space may be a strict subset due to cancellation, potentially missing important interference directions; the impact of projection threshold and rank selection is not fully discussed.
- Experiments are mainly based on the Stable Diffusion family with sequences of up to 100 edits; generalizability to longer sequences, newer architectures (e.g., FLUX), and training-based erasure methods is limited.
- On erase-set metrics, CEASE does not erase most deeply; the authors explain this as "redirecting to coherent anchor substitutes" rather than weak erasure, but there is no quantitative characterization of the trade-off between erasure thoroughness and generation coherence.
- In baseline comparisons, some methods (e.g., UCE, RECE) perform extremely poorly in the continual setting, which may exaggerate CEASE's relative advantage; comparison with stronger continual-learning baselines is insufficient.
Questions for authors
- After adding the anchor token to the invariance matrix in AIC, does it affect preservation of non-anchor prompts? How is βᵀc*≈0 guaranteed under numerical precision with a zero regularizer?
- How are the SVD rank and projection trigger condition in HOC set? Are they sensitive across tasks? Can an adaptive selection strategy be provided?
- The column space of cumulative update Δ_hist may lose historical directions; have you considered keeping explicit historical α vectors or using more stable subspace tracking?
- CEASE's CLIP score on the erase set is higher than some baselines; how do you distinguish "redirecting to the anchor" from "incomplete erasure"? Can human evaluation or an anchor semantic consistency metric be added?
- Is the method applicable to training-based erasure or non-closed-form updates? How do interference and computational overhead change for longer sequences (e.g., 500+)?
2 comments
Response to Reviewer 3hvk (Part 2/2)
W4 / Q4. CEASE does not erase most deeply… there is no quantitative characterization of the trade-off between erasure thoroughness and generation coherence. How do you distinguish "redirecting to the anchor" from "incomplete erasure"? Can human evaluation or an anchor semantic consistency metric be added?
As you suggested, we measured anchor consistency as the CLIP similarity between the SD1.4 erase-set images after 100 edits and the anchor "person":
| Method | CLIP similarity to anchor ↑ | CLIP similarity to original prompt |
|---|---|---|
| Original model | 17.10 | 30.95 |
| SPEED | 18.11 | 18.38 |
| CEASE | 19.88 | 21.54 |
CEASE's erased images are the closest to the anchor. They also stay more aligned with the original prompt than SPEED's, because the redirected images keep the non-identity content of the prompt, such as the portrait style or scene. Meanwhile, the Qwen2-VL-7B identity judge recognizes the erased celebrity in only 3.5% of these images. Following Reviewer UtoL's suggestion, we also computed LPIPS separately on the person region and the background using person masks: on the erase set, CEASE substantially changes the person region while changing the background less than SPEED, so its edits are confined to the identity. Together, this indicates that CEASE redirects targets to the anchor rather than leaving them incompletely erased. The erasure depth is also adjustable: a larger erasure strength scales each update for deeper erasure. On SD2.1 after 100 edits:
| Erasure strength | Target Acc ↓ | Retain Acc ↑ | COCO LPIPS ↓ | COCO FID ↓ |
|---|---|---|---|---|
| 1.0 | 34.2 | 69.6 | 0.290 | 38.20 |
| 1.5 | 12.0 | 68.6 | 0.310 | 40.78 |
| 2.0 | 3.4 | 66.6 | 0.339 | 41.77 |
Larger strengths progressively deepen erasure. We will add both analyses.
W3 / Q5. Generalizability to longer sequences, newer architectures (e.g., FLUX), and training-based erasure methods is limited. Is the method applicable to training-based erasure or non-closed-form updates? How do interference and computational overhead change for longer sequences (e.g., 500+)?
As you suggested, we extended the SD1.4 celebrity task to 500 sequential edits of distinct celebrities:
| Method | Target Acc ↓ | Retain Acc ↑ | COCO LPIPS ↓ | COCO FID ↓ |
|---|---|---|---|---|
| Reference | 71.7 | 72.0 | 0.000 | 0.0 |
| SPEED | 0.4 | 47.2 | 0.628 | 97.2 |
| CEASE | 0.7 | 57.2 | 0.596 | 83.3 |
Both methods keep target recognition very low, and CEASE still preserves retained identities and general images better, so its advantage persists over longer sequences. AIC and HOC add only about 2% to SPEED's edit time.
For training-based erasure, both constraints apply to any additive update: AIC keeps the anchor response unchanged, and HOC projects the update's outputs away from the selected historical directions. A training-based editor can therefore use them to restrict the range of its parameter updates.
For newer architectures, we evaluated 100 sequential edits on SDXL and FLUX.1-dev (DiT, flow matching). On FLUX.1-dev, we edit the text-stream value projections and evaluate 200 target and 200 retain images.
| Model | Method | Target Acc ↓ | Retain Acc ↑ | COCO LPIPS ↓ | COCO FID ↓ |
|---|---|---|---|---|---|
| SDXL | Reference | 88.8 | 72.6 | 0.000 | 0.0 |
| SDXL | SPEED | 5.4 | 65.0 | 0.345 | 39.1 |
| SDXL | CEASE | 4.4 | 65.6 | 0.327 | 36.8 |
| FLUX.1-dev | Reference | 63.0 | 68.0 | 0.000 | 0.0 |
| FLUX.1-dev | SPEED | 18.0 | 52.0 | 0.588 | 61.0 |
| FLUX.1-dev | CEASE | 19.5 | 53.0 | 0.509 | 54.1 |
On both backbones, CEASE keeps target and retain accuracy comparable to SPEED and changes general images less.
W5. Some methods (e.g., UCE, RECE) perform extremely poorly in the continual setting, which may exaggerate CEASE's relative advantage; comparison with stronger continual-learning baselines is insufficient.
Table 1 already compares CEASE with two continual unlearning baselines, CUIG-GradProj and Sculpting Memory, run from their public implementations. Both are training-based and constrain how each request is optimized, but not the weight update that is finally committed. CUIG-GradProj projects the gradient, yet the committed Adam update leaves the protected subspace, and this leakage grows by several orders of magnitude over the sequence. Sculpting Memory restricts which parameters each request updates, but does not prevent an update from changing the outputs of the shared anchor or of earlier edits. Interference therefore accumulates in both, whereas AIC and HOC constrain the committed update directly. We will add this analysis.
Thank you again for your valuable comments.
Response to Reviewer 3hvk (Part 1/2)
Thank you for your encouraging review and for recognizing the practical setting, the solid analysis and the training-free design. We address each point below.
The rank-one form is exact for a single target token and anchor, not an approximation. AIC adds c\* to the invariance matrix (Eq. 9), and with a zero regularizer the closed-form correction enforces Δic\*=0 as an equality, giving βi⊤c\*=0. The condition is thus exact in exact arithmetic, and floating-point error in the solve leaves only a residual at numerical precision, which does not restore the anchor and cross channels in any meaningful amount. We will distinguish the exact constraint from its numerical implementation.
AIC protects non-anchor prompts rather than restricting them. Because every edit redirects to the same anchor, any prompt with a component along c\* accumulates noise through the anchor and cross channels (Eq. 8). Once βi⊤c\*=0, both channels vanish for every prompt (Eq. 10). Consistently, CEASE improves retain-set LPIPS and all COCO metrics over SPEED on both SD1.4 and SD2.1.
In each layer, HOC triggers when the interference ∣Δhistc1(i)∣2, i.e., how much past updates already change the current target embedding, exceeds μ+ησ, where μ and σ are its running mean and standard deviation (EMA, decay 0.9). It then keeps the singular directions whose eigenvalue exceeds 1e-6 of the largest, up to a rank cap given as a fraction of the output dimension. The defaults are η = 1.5 and a 75% rank cap. We varied each one at a time on SD1.4 after 100 edits, with the erasure strength fixed at its default of 1.0:
Across all settings, Target Acc varies by at most 0.8 points, Retain Acc by at most 1.2 and COCO FID by at most 1.1. CEASE is therefore not sensitive to these hyperparameters. The same defaults work on the instance task without re-tuning, and we use them for all tasks and backbones.
As you suggested, we also compared the cumulative SVD with an explicit bank of per-update directions in a matched 200-edit SD1.4 run, using CLIP top-1 accuracy on 100 target and 100 retain celebrities:
The two variants perform comparably, so the explicit bank offers no clear advantage. At the endpoint, the cumulative update also preserves the full span of the stored directions in all examined layers.