acceptodds
Under review as a conference paper at ICLR 2027

AGTTU: Attention-Guided Target Token Unlearning for Large Language Models

Abstract

Large language models exhibit strong language understanding and reasoning capabilities, but also memorize private or copyrighted content, raising privacy and compliance concerns. Existing unlearning methods often uniformly suppress entire forget samples, inadvertently degrading general language capabilities unrelated to the unlearning targets and causing excessive forgetting. Token-level approaches alleviate this issue, but some rely on auxiliary models or additional perturbation-based scoring. We propose Attention-Guided Target Token Unlearning(AGTTU), which selects target tokens based on the attention they receive from subsequent answer positions, using the model's own attention allocation to guide targeted unlearning. We further introduce Contrastive Logits Preference Loss (CLPL) to suppress target tokens while limiting damage to non-target content, together with a retain-set objective to preserve model utility. Experiments on TOFU and MUSE Books demonstrate that AGTTU improves the trade-off between unlearning effectiveness and model utility. On TOFU with Llama-3.2-1B, AGTTU achieves higher reported forget quality and model utility than the listed baselines across all three forgetting ratios. On MUSE Books, compared with TPO, AGTTU reduces the verbatim memorization score on the forget set by 4.51% while increasing the knowledge memorization score on the retain set by 15.17%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Model Unlearning, Target Token Unlearning, Attention-Guided Unlearning, Selective Forgetting, Machine Unlearning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qYiBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AGTTU, an attention-guided target token unlearning framework for LLMs. The method leverages the model's internal attention scores to distinguish critical "target tokens" from "non-target tokens" within the forget set. It then introduces a Contrastive Logits Preference Loss (CLPL) to precisely suppress target tokens while penalizing the degradation of non-target tokens. Combined with a retention constraint, AGTTU achieves effective unlearning without compromising the model's general utility, outperforming baseline methods.

Critical strengths

1.Intuitive and Efficient Motivation: The idea of leveraging the model's inherent attention patterns to identify unlearning targets is both intuitive and computationally efficient. It avoids the need for auxiliary identifiers or external scoring models.

2.Fine-Grained Unlearning Strategy: By distinguishing between "target tokens" (to be forgotten) and "non-target tokens" (to be retained) within the forget set, the method enables more precise unlearning compared to coarse-grained approaches, effectively mitigating over-forgetting.

3.Well-Designed Loss Function: The proposed Contrastive Logits Preference Loss (CLPL) thoughtfully introduces a margin and a one-sided penalty to balance target suppression and non-target retention.

Critical weaknesses

1.Limited Generalizability: The evaluation is restricted to two extremely small-scale models (Llama-3.2-1B and 3B). The lack of experiments on mainstream larger models (e.g., 7B, 13B) raises serious concerns about the method's scalability and generalizability, as unlearning difficulty typically increases with model size.

2.Limited Technical Novelty: The method largely combines existing components. The attention-based target token selection closely resembles prior work like DTO, and the proposed CLPL loss is mathematically almost identical to TPO's LPL, as the authors themselves acknowledge. Overall, the contribution appears more compositional than a substantive methodological advance.

3.Insufficient Mechanism Analysis: The paper lacks a rigorous justification for equating "high attention" with "high unlearning value." High attention scores may simply correspond to syntactic connectors rather than core knowledge. A qualitative analysis of the selected tokens is needed to validate the fundamental assumption of the method.

Questions for authors

1.Could the authors provide a more detailed justification for the choice of model scale? Given that unlearning complexity often increases with model size, do the authors have any preliminary evidence or theoretical arguments suggesting AGTTU can maintain its performance on 7B or larger models? 2.To better understand the novelty, could the authors explicitly articulate the key technical barriers that were overcome when integrating attention-based selection with the CLPL loss? What specific failure modes of simply applying DTO or TPO independently does the joint framework resolve?

5 comments

W1 Q1:Evaluation is limited to Llama-3.2-1B and 3B, leaving the method’s scalability and generalizability to larger models, such as 7B and 13B, unverified.A more detailed justification of the chosen model sizes and preliminary empirical evidence or theoretical arguments are needed to support AGTTU’s performance on models with 7B parameters or more.

A1: We thank the reviewer for the follow-up question about model scale. We selected Llama-3.2-1B and Llama-3.2-3B to align with prior work: TPO evaluates Llama-3.2-3B and Llama-2-7B, while DTO evaluates Llama-3.2-1B and Llama-2-7B. Due to computational constraints, our initial 7B evaluation focused on MUSE-Books. We have now added experiments with Llama-2-7B on TOFU forget01/05/10 to the main experimental section. AGTTU’s results on Llama-2-7B are shown below: Table 3: Results on TOFU with Llama-2-7B.

MethodForget 1% FQ↑Forget 1% MU↑Forget 5% FQ↑Forget 5% MU↑Forget 10% FQ↑Forget 10% MU↑
Retain1.00000.62611.00000.62611.00000.6125
Original0.00130.62801.33×10⁻¹³0.62801.98×10⁻²⁵0.6280
GradDiff0.00680.57081.94×10⁻¹¹⁹0.56634.31×10⁻²²³0.5730
NPO0.76590.58160.11230.53150.28120.5756
TargetNPO0.91880.57840.22050.61380.01000.5843
TPO (ChatGPT)0.76590.61360.17790.41220.21070.5454
DTO0.76590.60020.14210.49660.13140.4874
AGTTU(Ours)0.76590.60180.86550.61240.58120.6192

AGTTU achieves substantially higher forgetting quality on both the 5% and 10% splits. These results address the concern that unlearning becomes more difficult as model size increases, showing no degradation in AGTTU’s effectiveness at the 7B scale. Together with the 7B results on MUSE-Books presented in the initial submission, these findings support AGTTU's effectiveness at this model scale.

W2 Q2: The method largely combines existing components, so its substantive novelty requires clarifying the technical barriers overcome by jointly designing attention-based selection and CLPL, and the specific failure modes of standalone DTO or TPO that this framework resolves.

A2: We thank the reviewer for this question. Integrating attention-based selection with CLPL addresses two specific technical challenges.

1. Substantially reducing token-selection time. Among the selectors we compare, attention is the only one that obtains a per-token influence signal with O(1) forward passes. A single prefill pass yields the column sums of the final-layer attention matrix: the sum of column i aggregates the attention paid to token i by all subsequent answer positions, serving as an estimate of its importance to generating the remaining answer. In contrast, loss-based selection and DTO require O(L) token-wise computations, where L is the answer length—approximately 37 forward passes per example on TOFU and 2,048 per chunk on MUSE. NER and TPO require external or auxiliary models, and TPO’s local identifier also requires masking each answer token individually. Thus, attention is the only selector in our comparison that is both self-contained and requires O(1) forward passes. We have added timing comparisons of local token selectors to Appendix E.4, using Llama-3.2-1B on TOFU forget10, as shown below.

SelectorSelection Time (s) ↓Uses an Auxiliary Model
NER (spaCy)24.4Yes
Loss-based184.1No
Gradient saliency19.6No
TPO (DistilBERT)104.5Yes
DTO (delta)34.8No
Attention (ours)8.8No

2. CLPL unifies target suppression and non-target protection into a single reference-based objective. The reference model is the original model before unlearning, θ0\theta_0. TPO’s LPL also protects non-target tokens in the forget set, but does not constrain them relative to reference values. CLPL expresses both terms using the logit difference from θ0\theta_0, ziθ0−ziθz_i^{\theta_0} - z_i^\theta, and applies different constraints to each: the target term introduces a margin mm, requiring each target logit to decrease by at least m relative to its reference value; the non-target term is one-sided, increasing a logit only when it falls below its reference value. In other words, CLPL measures each token’s logit deviation from the reference model but acts in opposite directions: target tokens are suppressed, while non-target tokens are pulled back up only when their logits fall below the reference values.

Thus, applying DTO or TPO alone leaves several issues that our joint framework addresses: (i) local token selection in DTO and TPO requires O(L) token-wise forward passes, with computational cost increasing with answer length; (ii) TPO’s external-model variant introduces potential privacy risks; and (iii) LPL does not constrain non-target logits relative to reference values. In contrast, CLPL unifies target suppression and non-target protection through each token’s logit difference from the reference model θ₀. The target term introduces a margin to enforce a larger decrease, while the non-target term applies a one-sided penalty only when the logit falls below its reference value, balancing effective forgetting with the preservation of general capabilities.

W3: A qualitative analysis of selected tokens is needed to establish whether high attention identifies core knowledge rather than syntactic connectors, supporting its use as a proxy for unlearning value.

We thank the reviewer for raising this point. Our rationale rests on a principled connection: work on KV-cache optimization has demonstrated that attention scores can identify tokens important for generation. H2O and SnapKV retain these tokens' key–value pairs while discarding others, preserving generation quality under cache compression. Unlearning, in turn, seeks to disrupt the generation of specific answers. The same importance signal thus informs which tokens to retain during inference and which to suppress during unlearning, suggesting a duality between the two objectives. We empirically tested this reverse application: masking attention-selected tokens reduces the generation probability of the full answer more than random masking, supporting the signal’s relevance to unlearning.

To address the concern that high attention may merely identify syntactic connectors, we analyzed the part-of-speech distribution of selected tokens. Content words (e.g., nouns and verbs) account for 68.4% of attention-selected tokens, compared with 59.7% across all answer tokens, while function words (e.g., prepositions and conjunctions) account for 27.1%, compared with 29.9% overall. Content words therefore dominate the selection, which is inconsistent with the hypothesis that attention merely selects syntactic connectors.

Reviewer_Wq5uOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AGTTU selects answer tokens for unlearning using the attention they receive from later answer positions in a frozen model. It combines this selector with CLPL, which encourages target-logit reduction and penalizes decreases at non-target positions, plus retain-set cross-entropy. Results on TOFU and MUSE Books show useful forgetting–utility trade-offs, and ablations examine token selection and the two additions to the loss.

Critical strengths
  • The selector is practically appealing. Target masks and reference logits are computed offline from the original model, without an auxiliary identifier or repeated tokenwise perturbation. The paper specifies the attention-to-prediction alignment and preserves the complete teacher-forced context, making the procedure reasonably clear to reproduce.

  • The matched token-selection experiment supports the main idea. Table 4 holds the loss and token count fixed, and high-attention selection gives higher FQ than random or low-attention selection at every tested ratio. I find this more persuasive than the masking motivation alone. The paper also reports the accompanying utility trade-offs rather than presenting the selector as uniformly better on both axes.

Critical weaknesses

The selection rule has a specific boundary case that can oppose the intended optimization. For a multi-token answer, the final token receives score zero and is normally assigned to the non-target set, whose logit decreases CLPL penalizes. If the forgotten entity, value, or an important entity subword occurs there, the procedure may explicitly protect part of the target fact. Appendix B acknowledges that this token can contain knowledge, but does not evaluate the consequence. The single-token exception does not address multi-token answers ending in factual content. This matters for interpreting the selector as a way to distinguish content to forget from content to preserve.

Questions for authors

See Weakness.

4 comments

Q1: The final token of a multi-token answer receives a zero attention score and is typically protected by CLPL as a non-target token, potentially hindering the forgetting of factual content at the end, yet this boundary case remains unevaluated.

A1: We thank the reviewer for this precise observation. We agree that this boundary case should be evaluated rather than merely acknowledged. We therefore quantified its frequency and conducted a direct ablation in the most relevant setting.

First, we clarify that the zero-score rule applies only to the absolute final token of the answer. The penultimate token still receives attention from the final token and is scored normally. Thus, this concern applies specifically when factual content occupies the final token. We examined this case on both datasets.

On TOFU, all 640 forget-set answers across forget01/05/10 end with a punctuation token (a period); none ends with a letter or digit. Thus, the zero-scored final token is always punctuation rather than an entity token, and this boundary case does not protect part of the target fact on TOFU.

On MUSE-Books, the forget set consists of raw book text segmented into fixed-length windows of 2,048 tokens. The final token therefore reflects an arbitrary chunk boundary rather than a content-defined boundary. Across 553 chunks, 78.84% of final tokens occur at boundaries between words or within truncated words; the remainder are punctuation or whitespace. Even when a final token belongs to an entity to be forgotten, the zero-score rule affects only one of the 2,048 tokens, while the entity's other tokens remain eligible for normal scoring and selection.

To directly assess the impact, we conducted an ablation that forcibly includes the final token of every chunk in the target set. We performed this test on MUSE-Books, where the concern is most relevant, since TOFU answers always end with punctuation rather than factual content. VerbMem, KnowMem (Df), KnowMem (Dr), and PrivLeak remained nearly unchanged, indicating that this boundary case has a negligible impact on the results in this setting.

We have revised the discussion in Appendix B and added the empirical statistics and ablation results described above.

MethodVerbMem↓KnowMem Df↓PrivLeak→0KnowMem Dr↑
AGTTU (final token forced into target set)0.6924.73-39.2956.49
AGTTU0.6923.46-34.4457.24

Dear Reviewer Wq5u,

We sincerely thank you for reviewing our paper and for helping us improve our work. With only a short time remaining in the discussion period, we would be very grateful to know whether our responses have addressed your concerns. In particular, we are eager to learn whether our latest response has resolved any remaining issues. We would be happy to continue the discussion if anything remains unclear.

Best regards,

Authors

Thanks for your rebuttal. My concerns have been resolved. I will maintain my score.

Dear Reviewer Wq5u,

Thank you very much for your positive assessment and for confirming that our response addressed your concerns. We greatly appreciate your time and feedback.

Best regards,

Authors

Reviewer_Y9PBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AGTTU, a token-level unlearning method for LLMs. This paper defines the targeted tokens to unlearn in the target forget-set answer by the attention they receive from subsequent answer positions in the last layer. Then the unlearning algorithm suppresses these selected targeted tokens with the constrastive logits preference loss, which is an extension of TPO's logit preference loss with a margin on the average target logit decrease and a one-side hinge penalty term on the non-targeted logits. Their experiments are conducted on TOFU datgaset and MUSE Books, demonstrating the utility and unlearning effectiveness trade-offs of the proposed method and related unlearning baselines.

Critical strengths

S1. The selection signal is cheap and self-contained. It needs a single prefill of the original model, with no auxiliary identifier model and no per-token perturbation scoring.

S2. The token-selection ablation (Table 4) compares strategies under a matched token budget, and the attention-masked NPO variant (Target NPO) is a useful control.

S3. The paper reports several unfavorable results openly (TOFU-3B at 1%, the CLPL ablation at 10%, KnowMem_f and PrivLeak on MUSE).

Critical weaknesses
  • W1: The method combines existing pieces with small modifications, so the novelty is quite limited. For example, scoring a token by the cumulative or average attention it receives from later positions has been used as a token-importance heuristic in prior KV-cache compression methods, such as H2O (Zhang et al., NeurIPS 2023) and SnapKV (Li et al., NeurIPS 2024). The paper does not discuss this line of work. Attention has also been explicitly leveraged for LLM unlearning, e.g., through attention smoothing in ASU (Zare Zade et al., ICLR 2026), which is cited but not empirically compared against.

  • W2: The necessity of attention-based selection is not demonstrated. There is also a conceptual gap: showing that a token influences the likelihood of later tokens is not the same as showing that it carries the knowledge to be forgotten. The paper gives no evidence for the second claim. It shows no examples of selected tokens and no analysis of overlap with entities or factual spans. It also does not analyze positional bias (earlier tokens are attended to by more positions) or attention-sink effects, which are known to dominate last-layer, head-averaged attention. More importantly, simple alternative selectors are missing.

  • W3: The paper compares attention selection only with Random and Low-p. Necessity requires comparison against cheap, local selectors such as a local NER or entity tagger, loss- or probability-based selection, gradient saliency, and entropy (ETW, Koh et al., 2026). A local selector would also address the paper's privacy objection to TPO's third-party identifier. That objection is itself incomplete: TPO also provides a local DistilBERT-based identifier and uses it on MUSE (Zhou et al., 2026). This existing local selector should be compared on TOFU, and the characterization in Sec. 2.2 should be corrected. Closely related token-level methods (SU, Wan et al., 2025; ETW; span-level unlearning, Yoon et al., 2026) and ASU are cited but not compared.

  • W4: The main tables mix results quoted from prior papers, baselines run by the authors, and AGTTU itself, and these entries were obtained under different evaluation pipelines, checkpoint rules, and data-access assumptions. The self-run baselines raise further concerns. The hyperparameters are missing. GradDiff collapses completely, and the TPO reproduction on Llama-3.2-3B falls far below the results Zhou et al. (2026) report for the same model, with no discussion. Together with the per-setting tuning of AGTTU's own hyperparameters and the missing original-model row, this makes the headline claim of outperforming all baselines hard to assess. All methods should be rerun in a single framework under identical settings.

Questions for authors
  • Q1. What evidence shows that attention-selected tokens actually carry the knowledge to be forgotten? Could you show examples of selected tokens and compare against simple local selectors (e.g., NER- or loss-based) under the same budget?
  • Q2. Since plain LPL outperforms full CLPL on both FQ and MU at 10% (Table 5), what evidence supports CLPL as a contribution? Could you report results over multiple seeds?
8 comments

W1: The paper offers limited novelty: H2O and SnapKV already use attention from subsequent positions to measure token importance, yet the paper neither discusses these works nor provides an empirical comparison with ASU.

We thank the reviewer for this constructive comment. We agree that attention-based token importance is not new. H2O and SnapKV demonstrate that attention-based importance can identify tokens or KV entries that are critical for preserving generation under a constrained KV budget. In the revised manuscript, we discuss these connections explicitly and clarify that we do not claim the underlying attention-importance heuristic itself as novel.

Our contribution concerns its reversed use in target-token unlearning. Specifically, we ask whether tokens that are important for preserving generation are also effective targets for suppressing it. Our masking comparison supports this reverse relationship: suppressing high-attention answer tokens reduces the probability of the complete answer more than masking equally many random tokens. Building on this observation, AGTTU selects target tokens with a single prefill of the frozen model, without auxiliary identifiers or repeated tokenwise perturbation forwards, and applies CLPL to suppress selected targets while explicitly limiting degradation of non-target answer content in the same reference-based objective. Under controlled comparisons, this design yields a better forgetting-utility trade-off than the evaluated target-selection strategies.

Regarding ASU, it smooths attention distributions and performs self-distillation at the loss level; it does not construct a target-token set and is therefore not included in our target-selector ablation. We now clarify this mechanistic distinction in the related work.We have updated the related work and contribution statement to acknowledge these connections and to state our novelty more precisely.

W2、Q1: whether attention-selected tokens carry the knowledge to be forgotten,requesting examples and comparisons with simple local selectors under the same budget.

A1: We agree with the reviewer’s distinction: a token influencing subsequent generation does not imply that it encodes the knowledge or entity to be forgotten. We do not claim that attention localizes entities or factual spans, and our forgetting objective does not require the selected tokens to be entity tokens. Under teacher forcing, unlearning aims to reduce the probability of generating the complete answer. We therefore select tokens whose suppression most effectively reduces the probability of the full answer—operationally important tokens—rather than tokens that are semantically identifiable as entities. This criterion is supported as follows:

  1. Masking attention-selected tokens reduces the probability of generating the complete original answer more than masking an equal number of random tokens, indicating that these tokens are important for answer generation.
  2. Under the same CLPL objective and equal target budgets, with top-p=0.3p=0.3 for 1% and 0.20.2 for 5%/10%,attention selection achieves higher forgetting quality than NER, loss-based, gradient-based, and DTO selectors(Table 1). These results address the necessity of attention-guided selection under a controlled comparison. Moreover, FQ is not a superficial measure of output corruption: in OpenUnlearning, FQ is the pp-value of a two-sample Kolmogorov–Smirnov test comparing the truth-ratio distribution of the unlearned model with that of the Retain model. A high FQ indicates no statistically significant distributional difference from the model trained without the forget data.

Table 1. Selector ablation study on TOFU using Llama-3.2-1B.

Method1% FQ↑1% MU↑5% FQ↑5% MU↑10% FQ↑10% MU↑
NER0.09710.58404.02×10⁻⁶0.56561.06×10⁻⁸0.5725
Loss-Based0.40460.56310.00010.58076.92×10⁻⁵0.5890
Gradient0.91880.57760.06800.57451.46×10⁻⁶0.6054
ChatGPT (TPO)0.16500.56331.10×10⁻¹⁴0.56613.91×10⁻⁸0.5707
Bert (TPO)0.01430.58173.08×10⁻¹²0.57185.80×10⁻¹⁴0.5721
DTO0.09710.58417.53×10⁻⁵0.57919.98×10⁻⁷0.5788
Attention (AGTTU)0.99000.55670.54530.57270.15460.5709

Attention (AGTTU) is our method; TPO uses all unwanted words, following the original paper.

  1. Among the tokens selected by attention, content words, such as nouns and verbs, account for 68.4%, compared with 59.7% in all answers, whereas function words, such as prepositions and conjunctions, account for 27.1%, compared with 29.9% in all answers.The selected tokens are therefore skewed toward substantive content rather than syntactic connectives.

We have added examples of tokens selected by different selectors in Appendix E.3 of the revised manuscript. Below are two concrete examples of attention-based target selection:

Example 1

Q: What specific genre is Nikolai Abilov known for?

A: Nikolai Abilov is most celebrated for his compelling writing in the African American genre, bringing fresh perspectives through his unique cultural lens.

Target tokens:genre, bringing, perspectives, unique, lens

Example 2

Q: What gender is author Basil Mahfouz Al-Kuwaiti?

A: Author Basil Mahfouz Al-Kuwaiti is male.

Target tokens:Author, male

We also analyzed the positional distributions of tokens selected by different selectors and added a line plot to Appendix E.3 of the revised manuscript to illustrate their relative positions within answers: attention-selected tokens are heavily concentrated toward the end of the answer (65.5% in the final quartile versus only 3.9% in the first), whereas NER and TPO exhibit strong biases toward the beginning (71% and 81% in the first quartile, respectively), ruling out the possibility that attention merely selects early positions or attention sinks.

W3: Comparisons with inexpensive local selectors, including TPO’s local DistilBERT identifier, and related unlearning methods are needed, along with a correction of the incomplete discussion of TPO's privacy implications in Sec. 2.2.

We thank the reviewer for this accurate and thorough observation. We have expanded the token-selection ablation beyond Random/Low-p to include inexpensive local selectors—NER, loss-based selection, gradient saliency, TPO, and DTO (Table 1)—and added this ablation to Appendix E.1. Attention achieves the highest forgetting quality on every split, indicating that its advantage is not merely a consequence of the previously limited comparisons.

We have corrected Sec. 2.2 in the revised manuscript: in addition to its GPT-based identifier, TPO indeed provides a local DistilBERT-based identifier and uses it on MUSE. For our MUSE unlearning experiments with Llama-2-7B, the TPO results we compare against are those reported in the TPO paper, which use its DistilBERT-based identifier.

We position our work as token-level unlearning and select DTO and TPO as baselines because they are recent token-level unlearning methods that we consider most relevant and use the same TOFU FQ/MU evaluation protocol, allowing direct comparison. Code for SU and ETW is not publicly available, while Span-Level uses different datasets and evaluation metrics. ASU uses different TOFU metrics but the same MUSE metrics as ours. We therefore compare ASU’s reported results with AGTTU on MUSE-Books in the table below, showing that AGTTU remains competitive.

Table 2

MethodVerbMem↓KnowMem Df↓PrivLeak→0KnowMem Dr↑
ASU (GD)4.919.0-52.358.9
ASU (KL)5.328.6-51.062.5
AGTTU (Ours)0.6923.46-34.4457.24

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.