- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes TASE enabling a very sparse method for LLM unlearning. The authors claim that they observe that editing only a few token embeddings can change a model’s outputs on the forget set. TASE tries to pick a few “anchor” tokens that are common in the forget set but rare in the retain set. It then applies gradient ascent only to the input embeddings of these tokens and keeps everything else frozen. They claim that editing needs no retain-set gradients and no reference model. The experiments show forgetting and retained utility are competitive with or better than the other baselines.
Critical strengths
The paper has several highlights.
The observation is simple and interesting. Editing only a few input embeddings is enough to strongly suppress target outputs in a 7B model.
The method is cost-efficient. It needs neither retain-set optimization nor a reference model. It also uses less GPU memory and time than the baselines.
The experiments have broad coverage. They include encoder and decoder models, three to four model families, and both Q&A and long-document unlearning. The paper also reports extra checks such as MMLU, WMDP-Cyber, prompt robustness, recovery attacks, membership inference and ablations.
Critical weaknesses
Several concerns about the paper.
First, the core assumption is a bit questionable. So far as we know, factual knowledge in LLMs is mainly stored in the MLP and attention layers, not in input embeddings. Editing the embeddings of anchor tokens may only stop the model from recognizing the entity in the prompt. Maybe the knowledge itself still remain in the weights.
Second, editing a token embedding affects every input that contains this token. This can be dangerous as other entities that share the same name fragments would also be affected. Unfortunately, retain-set metrics and MMLU cannot detect this damage.
Als, the results are not consistent across benchmarks. TASE works very well on TOFU, however, it is a bit weaker on other settings, such as Phi-1.5B forget10. The method may depend on benchmark properties.
Questions for authors
-
Regarding the concern in the Weaknesses part, does TASE really remove knowledge, or only block recognition of the entity in the input?
-
How does TASE perform when the target knowledge has no unique tokens? For example, "real people with common names".
-
What happens to unrelated inputs that contain the anchor tokens?
-
For evaluation, could you compare with LoRA-based and any other localized unlearning methods (such as the following listed ones)? -SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation -Who's Harry Potter? Approximate Unlearning in LLMs ...
5 comments
How does TASE perform when the target knowledge has no unique tokens? For example, "real people with common names".
TASE's anchor selection relies on relative Forget–Retain frequency contrasts at the subword level, rather than requiring an exclusive whole-name token. Shared name fragments and other eligible subwords can still provide useful selection signals when their occurrence frequencies differ between Forget and Retain data.
Distinctiveness analysis. For each entity in TOFU, its 20 QA examples were compared against those of the other entities. D@5 measures the mean log(1 + frequency ratio) of the five highest-scoring eligible tokens. Based on anchor distinctiveness, we divide all entities into three groups:
| Distinctiveness group | Entity count | Unique QA Coverage |
|---|---|---|
| Low | 67 | 56.3% |
| Medium | 66 | 88.2% |
| High | 67 | 95.1% |
Here, Unique QA Coverage measures the proportion of an entity's QA examples containing at least one eligible token absent from the other 199 entities. Then, we constructed matched Low- and High-distinctiveness forgetting requests from the standard forget10 author pool. Each request contains six authors (120 Forget QA), with the remaining 194 authors (3,880 QA) serving as Retain data. Both requests share the same initialization. The Low request has 68.3% Unique QA Coverage, compared with 100% for High. Results are reported as mean ± sample SD:
| Group | Forget Coverage ↑ | Retain Exposure ↓ | FTR ↑ | F-RG ↓ | R-RG ↑ | MU ↑ |
|---|---|---|---|---|---|---|
| Low | 96.1% ± 2.7% | 24.6% ± 1.8% | 0.709 ± 0.017 | 0.306 ± 0.073 | 0.896 ± 0.022 | 0.601 ± 0.006 |
| High | 100.0% ± 0.0% | 13.3% ± 0.2% | 0.719 ± 0.034 | 0.253 ± 0.081 | 0.980 ± 0.005 | 0.613 ± 0.002 |
TASE remains effective as anchor distinctiveness decreases. In the Low group, where nearly one-third of the Forget QA contain no unique eligible token, TASE still achieves 96.1% Forget Coverage. Compared with the High group, lower distinctiveness increases Retain Exposure from 13.3% to 24.6% and moderately reduces R-RG from 0.980 to 0.896, while FTR and overall model utility remain close. These results show that relative subword-frequency contrasts continue to provide effective localization signals even when unique-token availability is substantially reduced.
QA-level zero-unique analysis. We further examined TOFU's token characteristics together with its unlearning behavior. Among the 640 standard-split Forget QA examples, 65 (10.2%) contain no unique eligible token relative to their designated standard-split Retain vocabulary. Thirty of these examples have corresponding controlled per-sample outputs, meaning that each QA was individually evaluated using outputs generated by the edited model under the same controlled Low/High setting. These zero-unique QA exhibit higher residual target-answer probability and ROUGE than the non-zero-unique examples in the same evaluation. Despite this increased difficulty, the Low-distinctiveness request containing such examples still achieves high Forget Coverage together with 0.896 R-RG and 0.601 MU. The results therefore show that TASE continues to achieve a favorable forgetting–retention balance when the forgetting request includes QA examples without unique eligible tokens.
Together, these results show that TASE does not rely on exclusive entity-name tokens. Relative subword-frequency differences remain sufficient to identify effective anchors under reduced token distinctiveness, enabling broad target coverage while preserving retained knowledge for shared-name and common-name targets.
What happens to unrelated inputs that contain the anchor tokens?
Unrelated inputs that contain edited anchor tokens can experience some degradation because they use the same modified embedding rows. TASE limits this effect by reducing Retain Exposure during anchor selection and constraining the embedding updates, allowing most retained knowledge to remain well preserved.
Matched collision–control analysis. We evaluated 153 anchor-exposed Retain QA and matched unexposed controls across 31 retained-entity clusters. For each pair, we compared the model change from Base to TASE and report the difference between exposed and control groups using 10,000 cluster-bootstrap resamples.
| Metric | Difference-in-differences | 95% CI |
|---|---|---|
| Exact Match | −0.719 | [−0.878, −0.462] |
| ROUGE-L | −0.464 | [−0.732, −0.204] |
| Gold-answer probability | −0.576 | [−0.784, −0.351] |
| Gold-answer NLL | +4.201 | [+1.549, +7.132] |
The matched comparison shows that shared-anchor exposure reduces retained-answer recovery and Gold-answer probability while increasing answer NLL. Thus, the retention effect is concentrated on the subset of Retain inputs that actually contain edited anchors.
Per-sample retain analysis. We further examined 2,400 retained-example evaluations corresponding to 800 unique Retain QA from the existing anchor-distinctiveness experiments (Q2). Each evaluation was grouped according to whether the input contained at least one selected and edited anchor.
| Retain evaluation group | Evaluation records | Gold-answer probability ↑ | Gold-answer NLL ↓ | ROUGE-L ↑ |
|---|---|---|---|---|
| Anchor-hit | 453 | 0.721 | 1.991 | 0.747 |
| Anchor-free | 1,947 | 0.986 | 0.025 | 0.982 |
Anchor-hit examples show lower retained-answer performance, whereas the much larger Anchor-free group preserves its original answers well. Retained-answer quality also decreases as anchor exposure increases.
As a result, although shared-anchor inputs can experience local degradation, TASE keeps such exposure limited and retains strong overall Retain performance. The selector controls which retained examples are exposed, while the constrained update controls how strongly the selected embedding rows are modified, together preserving most unrelated knowledge while maintaining effective forgetting.
Regarding the concern in the Weaknesses part, does TASE really remove knowledge, or only block recognition of the entity in the input?
TASE reduces the influence of targeted information on generated outputs through sparse edits to token embedding rows. These edits propagate through the fixed Transformer layers, changing contextualized representations involved in answer generation.
Representation Propagation (Appendix A.4.5, Figure 5). We traced the effects of embedding edits across Transformer layers. The representation drift progressively increases in both anchor and answer regions, while the first-order analysis captures the induced changes with high fidelity at moderate perturbation scales. This indicates that localized embedding edits induce changes that propagate through the Transformer and influence downstream answer representations.
Cross-prompt Robustness (Appendix A.4.4, Table 11). We evaluated target-answer recovery across eight prompt formulations, including paraphrases, aliases, multilingual queries, adversarial extraction, keyword-free inputs, and compositional reasoning. Target-answer probabilities remain low across all formulations. Although adversarial extraction, keyword-free, and compositional queries exhibit some residual lexical overlap with target answers, their target-answer probabilities remain low, indicating that TASE maintains effective suppression of target-answer generation even under these indirect formulations. These results demonstrate that TASE's suppression extends beyond the original entity wording to alternative formulations of the targeted knowledge.
Together, these results show that TASE induces downstream representation changes and reduces target-information reproduction across different query formulations, supporting representation-mediated unlearning rather than input-string-specific recognition blocking.