acceptodds
Under review as a conference paper at ICLR 2027

Linguistic Loopholes in LLM Unlearning: From a 174-Language–Script Benchmark to Coverage-Aware Unlearning

Abstract

Unlearning a fact in one language does not guarantee its removal in others as changing the query or even the requested answer language can reopen seemingly forgotten knowledge—a cross-lingual loophole. The most straightforward solution to this challenge – unlearning in all languages – is neither scalable nor desirable as it amplifies damage to unrelated model capabilities. We introduce the task of language budgeted multilingual unlearning with the goal of selecting a subset of languages that maximize cross-lingual erasure. To study this task we introduce Cross-lingual Unlearning Tensor, an unlearning benchmark that spans 174 language–script pairs and 25 atomic paraphrase types to examine when forgetting generalizes across linguistic expressions of the same knowledge. We further propose COVER, which selects source languages to maximize predicted COVERage of languages receiving no forget supervision; enabling unlearning on a language budget. Surprisingly, we find naively selecting strong individual sources does not reliably compose into strong source sets motivating our development of COVER. At deployment COVER only requires benign calibration data and access to the frozen model. Across three model families and two disjoint forget sets, COVER reduces mean held-out residual access by 7.8–27.3% relative to uniform source selection. We find these gains extend beyond synthetic benchmarks to real news documents in low-resource languages settings using human translated data from the Low Resource Languages for Emergent Incidents (LORELEI) corpus.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
llms, multilingual, unlearning, low resource languages

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

This paper studies the problem of cross-lingual loophole in LLM unlearning and provides a large-scale systematic analysis of when forgotten knowledge remains accessible across languages, scripts, and paraphrased expressions. The authors introduce the Cross-lingual Unlearning Tensor, covering 174 language-script pairs and 25 paraphrase types, and use it to analyze unlearning across different query and answer routes. They further define language-budgeted multilingual unlearning and propose COVER, a source-language selection method based on cross-language coverage estimated from benign calibration data. COVER improves held-out forgetting in several settings, although the gains vary across models and datasets and are more limited on the real-world LORELEI evaluation. Experiments on Aya, Qwen, and Llama, together with analyses of learning and unlearning transfer, language bottlenecks, script and paraphrase sensitivity, and source-set stability, show that cross-lingual forgetting varies greatly across linguistic routes and that effective multilingual unlearning depends on carefully selecting a limited set of source languages.

Critical strengths
  • The paper introduces a broad and well-structured benchmark for cross-lingual unlearning. The Cross-lingual Unlearning Tensor covers 174 language-script pairs and 25 paraphrase types, while separately evaluating query and answer language routes. This provides a much more systematic basis for studying cross-lingual forgetting than prior evaluations on a small number of languages.

  • The paper provides a systematic study of language-budgeted multilingual unlearning. By variously comparing source-language subsets across different models, language inventories, training orders, and disjoint forget requests, the experiments clearly show that source-language choice has a substantial and partly stable effect on held-out forgetting. This gives strong support for the need to carefully select a limited set of source languages.

Critical weaknesses

The limited advantage of COVER over simpler selection strategies is a central weakness of the paper. On TOFU, COVER selects the same source subsets as the fixed coverage rule for Aya and Qwen, indicating that it adds little benefit over a simpler heuristic. For Llama, COVER differs from fixed coverage, but its gains are mixed and it does not consistently outperform alternatives such as the historical-best subset. More importantly, on the human-translated LORELEI news data, COVER again coincides with fixed coverage and delivers only modest improvements, leaving limited evidence that the method provides a substantial advantage on more realistic data.

Questions for authors
  • Why does the learned COVER selector often choose exactly the same source subsets as fixed coverage, and what information learned from historical unlearning outcomes actually changes the ranking when the two methods differ?

  • Why are COVER’s gains substantially smaller on the human-translated LORELEI news data than on the synthetic benchmarks, especially for document-interior prompts? Could the authors provide further analysis of what causes this degradation in the more realistic setting?

4 comments
AuthorsOpenReview ↗

1. Additional baselines (Response to Reviewer xjZW)

We added geographic coverage, cosine-similarity coverage, English constrained coverage, and script diversity. COVER achieves 32.13 residual access versus 36.84 for the strongest added baseline on LUME. 3. Simple baselines presents the comparisons, including retention tradeoffs, alongside the existing baselines. Please see 4. Simple baselines. We have added the four baselines to Table 4b and their definitions to Appendix D.6 (“Additional source selection baselines”).

AuthorsOpenReview ↗

2. What information learned from historical unlearning outcomes actually changes the ranking (Response to Reviewer xjZW)

On our LUME forget request experiment, learning changes the selected source set and reduces residual access from 39.26 to 32.13, ranking first in both realizations. 4. What from historical unlearning outcomes actually changes the ranking when the two methods differ decomposes the exact scoring contributions responsible for that decision. Please see 4. What from historical unlearning outcomes actually changes the ranking when the two methods differ for the concrete examples and feature decomposition. The learned ranker is described in Section 4.1.

AuthorsOpenReview ↗

3. Smaller LORELEI/Interior gain (Response to Reviewer xjZW)

We investigated two aspects of the smaller LORELEI gains. (1) whether forgetting generalizes across document segments, and (2) how much improvement is available through source selection.

For document interiors, our stopping protocol leaves an additional generalization challenge that we did not sufficiently detail. In our experiment, unlearning stops based on source language forgetting at document opening, whereas interior evaluation uses different contexts and continuation targets within those documents. Across all 20 source triples and two training orders, mean oracle normalized source removal reaches 95.7% for Aya and 95.6% for Qwen on leads, compared with 86.9% and 89.0% on interiors. Thus, comparable forgetting at document openings does not imply comparable forgetting throughout the document, even in the directly supervised languages.

To put the smaller absolute selection gains in context, we also compared COVER with the best source triple identified in hindsight (via repeated empirical rankings). The best triple across both orders improves over uniform selection by 6.96 points for Aya and 3.91 for Qwen. COVER achieves 4.60 and 2.31 points, capturing 66% and 59% of that improvement. COVER therefore remains competitive among source selections on leads, while its benefit transfers less reliably to document interiors.

Reviewer_xjZWOpenReview ↗

Response to author

Thank you for your detailed response to my concerns. I will keep my positive rating.

Reviewer_DEh3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper focuses on multilingual LLM unlearning, where a fact removed in one language should also be inaccessible when the model is queried in other languages. The challenge is that forgetting in the supervised language does not transfer well to other query and answer languages. Existing studies of this cross-lingual gap cover only a few, mostly high-resource languages. To address this limitation, the authors build the Cross-lingual Unlearning Tensor, a benchmark based on TOFU and LUME that covers 174 language-script pairs and 25 paraphrase types. They then define a language-budgeted setting in which forget data is provided in only K source languages and propose COVER, which selects the source set using a learned ranker based on hidden-state similarity features computed on benign calibration data. Across three model families, COVER reduces held-out residual access by 7.8–27.3% relative to uniform source selection.

Critical strengths

The problem is real and important, and the benchmark is a valuable contribution.

The evaluation protocol is carefully designed, and the code for the Cross-lingual Unlearning Tensor is provided.

Critical weaknesses
  1. The motivation and evidence for COVER are weak. (1) COVER is tested only on three six-language inventories with K = 3 (20 candidate triples). The candidate space is very small, which does not effectively demonstrate the value of the language selection approach. As a result, the benchmark and COVER appear to be two independent studies. (2) It is unclear how much the learned ranker contributes. In Table 4(a), COVER is listed as "Fixed" for Aya and Qwen, matching the fixed-coverage selection. (3) Simple selectors are missing, such as similarity-based selection or a fixed set of representative languages chosen by consensus. (4) The paper states that "Gibberish scores remain low", but these scores are higher than those of the controls for Aya and Qwen (Table 5). Clearer evidence is needed to demonstrate COVER's effectiveness and generalization.

  2. The finding that forgetting does not transfer well across languages is already known from prior work [1, 2], so the main new contribution appears to be the benchmark's scale. The technical contribution and novelty therefore need further clarification. Moreover, given that this is primarily a benchmark paper, the evaluated baselines are insufficient.

[1] Lynch, Aengus, et al. “Eight methods to evaluate robust unlearning in llms.” arXiv preprint arXiv:2402.16835 (2024).

[2] Wang, Jiaan, et al. “Cross-lingual knowledge editing in large language models.” Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Questions for authors
  1. Why is unlearning in all languages neither scalable nor desirable? Could the authors demonstrate the difficulty of unlearning using all languages (translated from the forget query), for example, by comparing the unlearning costs of this baseline and COVER?

  2. Why is the language budget K the actual constraint? The motivation for the language-budgeted setting needs further clarification. Translating a forget request is inexpensive, while COVER itself requires aligned benign data in every candidate language. Could the authors explain this trade-off in more detail?

  3. What is the difference between naively selected sources and those selected by COVER? Does the paper provide a detailed comparison between these selections, such as showing detailed examples?

  4. Residual access is only meaningful if the model initially learned the fact in that language, but this is not the case for many of the evaluated languages. Why does the criterion for sufficient access require only a gain greater than 0.05 (Table 2)?

  5. Which unlearning method is used in Table 4(b) and 4(c)? Does COVER also help with other unlearning methods? Why were only SimNPO, RMU, and GradDiff selected?

Regarding the presentation:

  • In Table 2, what does "l" denote?

  • The presentation of the results is confusing, and the table and figure captions should be self-contained. It would also be better to report both the baseline (reference) and COVER results, as reporting only “+xxx” is confusing (Table 4c).

11 comments
AuthorsOpenReview ↗

1. The motivation and evidence for COVER are weak / What's wrong with all language unlearning / motivation for K as a constraint? (Response to Reviewer DEh3)

We include a full argument for this response in top level comments 1–3 (1, 2, 3) and summarize those responses here:

Unlearning in english only leaves substantial cross-language leakage in our experiments while unlearning all languages:

  • Has a utility tradeoff. Our results show greater gibberish output and code switching as a result of all language unlearning
  • Obtaining a full parallel corpus of your forget set across many languages may not be scalable.
  • LLM translations may perform poorly for unlearning human generated memorized facts

COVER can use existing public benign text for calibration, while forget supervision must address each new deletion request. 1. Why constrain K, what's wrong with all language supervision, motivation for budgeted multilingual unlearning develops this argument and finds that inexpensive generated translations provide less effective supervision than expensive human translation in our LORELEI experiment. We also note scenarios in which generating cheap translations may not be possible.

Please see top level comments 1–3 above. The corresponding additions in our revision appear in Appendix D.1 (Figure 8 and Table 23), Appendix D.3 (Table 25), and Table 4b, with the new baseline definitions in Appendix D.6.

AuthorsOpenReview ↗

2. COVER is tested only on three six-language inventories (Response to Reviewer DEh3)

We include new results with an expanded source pool from six to ten languages and increased the selection budget to K=4, increasing the number of possible subsets from 20 to 210. We increased our language evaluation pool to 25 languages, including 15 languages the checkpoint was not directly supervised on. COVER archives held out residual access of 0.179 compared with 0.256 for fixed COVER and 0.280 for individual top 4 (which selects the four languages that produce the strongest cross language forgetting when tested individually) and 0.427 for the mean of four random selections. These results extend our evaluation to a larger candidate pool and broader held out coverage. Please see 2. Larger language sets and new K values. These results appear in our revision in Appendix D.3, Table 25, summarized in Section 4.3.

AuthorsOpenReview ↗

3. What the learned ranker contributes (Response to Reviewer DEh3)

On our LUME forget request experiment, learning changes the selected source set and reduces residual access from 39.26 to 32.13, ranking first in both realizations. 4. What from historical unlearning outcomes actually changes the ranking when the two methods differ decomposes the exact scoring contributions responsible for that decision and explains the settings where learned and Fixed coverage coincide. Please see 4. What from historical unlearning outcomes actually changes the ranking when the two methods differ for the concrete examples and feature decomposition.

Reviewer_8EM4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a multilingual unlearning benchmark covering a large number of language,script pairs, studying how changing question language, answer language and phrasing affects extractability of apparently forgotten information. It proposes COVER to select languages for unlearning under a constraint on the number of source languages, aiming to improve forgetting across languages. The paper conducts experiments across multiple models and tries to measure forgetting and retained performance

Critical strengths
  1. The paper covers a large number of languages and scripts, and separately studies changes in question language, answer language and phrasing. I appreciate the breadth of this evaluation, and feel that this would be useful for understanding where apparent forgetting still leaves the same information accessible. I think the contribution of the multilingual benchmark would be useful for the community.
  2. The experiments have been designed well and with care, particularly matching how much forgetting has happened in the source languages before comparing how well it transfers. The comparisons of language combinations, and checking whether useful choices carry over to different forget requests, are also interesting. In addition to teacher forced likelihood/extractive evals the use of generative evals and the human-translated document experiments help strengthen the results
Critical weaknesses
  1. My primary concern is that I do not understand why the number of languages in the forget set would be a constraint. This is a central part of the paper, but I am not sure why this is a meaningful setting to study. I can understand limiting the size of the forget set or the optimization budget for unlearning, but I am less convinced by fixing the number of languages as the main constraint. Unlearning benchmarks such as RWKU [3] use synthetically generated forget sets, and there is related work on selecting examples or tokens for unlearning [1,2,4]. It would be interesting to use ideas from these methods to compare against selecting fewer examples across more languages under a comparable budget. Synthetic generation/rephrasing has already been used to improve pretraining data [5], so I would like to understand why translation should be a major cost here, given that rephrasing works at pretraining scale.

  2. The benchmark shows that changing the question or answer language can recover apparently forgotten information, but the main evaluation of COVER measures teacher-forced likelihood of the answer in the same language as the question. I think measuring, for each fact, the max teacher forced likelihood over the tested QA language/script combinations would be far more convincing when making claims that crosslingual lack of forgetting is addressed.

[1] Soumyadeep Pal, Changsheng Wang, James Diffenderfer, Bhavya Kailkhura, and Sijia Liu. Llm unlearning reveals a stronger-than-expected coreset effect in current benchmarks. arXiv preprint arXiv:2504.10185

[2] Yixin Wan, Anil Ramakrishna, Kai-Wei Chang, Volkan Cevher, and Rahul Gupta. Not every token needs forgetting: Selective unlearning to limit change in utility in large language model unlearning. arXiv preprint arXiv:2506.00876,

[3] Jin, Zhuo-Ran et al. “RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models.” ArXiv abs/2406.10890 (2024)

[4] Allouah, Youssef, Rachid Guerraoui, and Sanmi Koyejo. "Distributional machine unlearning via selective data removal." International Conference on Learning Representations. Vol. 2026. 2026.

[5] Maini, Pratyush, et al. "Rephrasing the web: A recipe for compute and data-efficient language modeling." Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Questions for authors
  1. How well does the language-selection procedure work when choosing from a much larger pool, say 50 languages rather than six? Does the selected set still provide good coverage of the remaining languages?
  2. The benchmark shows that changing the question or answer language can recover apparently forgotten information. Do the improvements from COVER still hold when the question and answer are in different languages? Could the authors evaluate these combinations, at least within the languages already used for COVER,
  3. What about using synthetically generated forget sets, as in RWKU? How would selecting a few languages compare with using fewer such examples across more languages, under a comparable data and optimization budget?
5 comments
AuthorsOpenReview ↗

1. Why the number of languages in the forget set should be a constraint (Response to Reviewer 8EM4 )

In 1. Why constrain K, what's wrong with all language supervision, motivation for budgeted multilingual unlearning we develop this argument through public benign calibration, the all-language generation comparison, and a direct test of automatically translated versus human-translated forget supervision. We have added supporting evidence in Appendix D.1, Figure 8 and Table 23.

AuthorsOpenReview ↗

2. Compare against selecting fewer examples across more languages under a comparable budget (Response to Reviewer 8EM4)

We thank reviewer 8EM4 for this suggestion to look at fixed forget set sample sizes with COVER selected languages versus all languages. We have added this comparison and it further strengthens the claims of the paper. We evaluate Aya, Qwen and Llama on two language sets. Each arm used the same forget facts with one fixed language version per fact. This. COVER distributes the examples across three selected languages, while all languages distribute them across 6. Forgetting measures oracle normalized mean reduction in answer probability. Including all languages narrowly beats out cover from the forget point of view, but this comex at utility cost. Combined with the reasons discussed in 1. Why constrain K, what's wrong with all language supervision, motivation for budgeted multilingual unlearning, this presents a clear tradeoff for why K is a worthwhile consideration.

TOFU

ArmForgotten fact recovery ↓[Gibberish] ↓
COVER (K=3)35.07%5.86%
All634.67%11.43%

These are averages across six model language contexts.

We also ran this comparison on LORELEI, where the model learned 20 news documents from human-written translations across six languages (Qwen3-4B and Tiny-Aya, two seeds). Under the same supervision budget, each document is supervised in three language versions, either the three COVER-selected languages or three of the six languages varying by document, and unlearning stops at matched source forgetting. Because TOFU's translations are themselves GPT-generated, we additionally test the all-language arm with Qwen3.5-27B translations in place of the human text, the inexpensive way to obtain every language. All arms are evaluated on the human text across eight evaluation languages.

**LORELEI**

ArmSupervision textEight-language forgetting progress ↑Retain answer probability ↑
COVER (K=3)Human translations50.0%67.8%
All (K=6)Human translations54.8%68.6%
All (K=6)Qwen3.5-27B translations19.5%68.7%

These are averages across Qwen3-4B and Tiny-Aya (two seeds each); retain is answer probability on retained documents (parent: 70.9%).

With human-written supervision, COVER again achieves 91.3% of all-language forgetting progress (50.0% vs 54.8%), consistent with TOFU. When the additional languages are instead produced by LLM translation, all-language forgetting of the human text drops to 19.5%, less than half of what COVER achieves with human text in only three languages, at the same retain probability. Translations are cheap, but they do not provide equivalently rich unlearning supervision (see 1. Why constrain K, what's wrong with all language supervision, motivation for budgeted multilingual unlearning).

We have added this comparison in Appendix D.1, Table 22, and summarize it in Section 4.3.

AuthorsOpenReview ↗

3. How well does the language-selection procedure work when choosing from a much larger pool? (Response to Reviewer 8EM4)

We include new results with an expanded source pool from six to ten languages and increased the selection budget to K=4, increasing the number of possible subsets from 20 to 210. We increased our language evaluation pool to 25 languages, including 15 languages the checkpoint was not directly supervised on. COVER archives held out residual access of 0.179 compared with 0.256 for fixed COVER and 0.280 for individual top 4 (which selects the four languages that produce the strongest cross language forgetting when tested individually) and 0.427 for the mean of four random selections. These results extend our evaluation to a larger candidate pool and broader held out coverage. Please see 2. Larger language sets and new K values. These results appear in our revision in Appendix D.3, Table 25, summarized in Section 4.3.

Comments on the paper

6 comments
AuthorsOpenReview ↗

4. What from historical unlearning outcomes actually changes the ranking when the two methods differ (DEh3, xjZW)

The clearest benefit appears on LUME experiment. Here learned COVER changes the selected sources from Japanese/Spanish/Chinese to English/Japanese/Russian, reduces held-out residual access from 39.26 to 32.13 (oracle normalized residual access, lower is better), and ranks first among all 20 triples in both realizations. Fixed COVER makes its selection since those languages have higher mean coverage - each mean held out languages strongest source match is better on average. Learned COVER does not make that selection for the following reasons:

QuantityInterpretationContribution favoring COVER’s set
Mean coverageAverage of each target language’s strongest source match−0.5476
Minimum coverageThe weakest target language’s strongest source match−0.5271
Backup coverageAverage strength of each target’s two strongest source matches+0.4564
Coverage variationHow much strongest-match scores vary across target languages+0.4299
Mean connection strengthAverage match across all selected source–target pairs+0.2702
Total score difference+0.0818

The positive contributions favor English/Japanese/Russian and negative contributions favor Japanese/Spanish/Chinese. Thus, COVER accepts weaker mean and minimum coverage because the other three terms jointly outweigh those disadvantages under the weights learned from historical unlearning experiments. Finally, evaluating both methods, we see residual access decrease from 18.71 to 14.84 (oracle normalized residual access, lower is better) from fixed to learned COVER.

As a second example, LRL Llama, learning replaces Nepali with Hindi, while keeping Arabic/Hausa fixed.

AuthorsOpenReview ↗

3. Simple baselines (DEh3, xjZW)

We agree there are some additional intuitive baselines that should be included. We have added some in the appendix (D) of our revised paper. We summarize the results below, and also remind the reviewers of our existing baselines already reported in the paper.

Existing baselines
Our paper already evaluates four source selection baselines against COVER. These are presented in Tables 4-5.
1. Uniform Selection: Averages all 20 triples
2. The historical best: Reusing the strongest language set from a previous deletion request
3. Top-3: The three strongest individual choices, which test if individually effective languages from an effective combination
4. Fixed COVER: Our COVER method without the learned weights

We summarize the advantages we see from COVER against these baselines: On LUME COVER achieves held-out residual 32.13 versus 39.26 for Fixed, 38.82 for Historical best, and 40.84 for uniform, ranking first among all 20 triples in both realizations.

In generated answer evaluation, COVER reduces full leakage relative to Individual top-3 for Aya (29.76% to 25.57%) and Llama (32.37% to 25.72%), and tying Qwen.

New Baselines

We have expanded our evaluation to include 4 additional baselines: cosine similarity COVER without learned weights, English constrained COVER, script/family heuristics, geographic typological coverage. We have added these four baselines to Table 4b, with their definitions in Appendix D.6 and a brief discussion in Section 4.3.

The results of COVER versus these baselines are summarized below on LUME dataset (update to table 4, b):

SelectorHeld-out residual ↓Extra retain damage ↓
Geographic coverage: strongest tested typology variant here36.84-0.62
Cosine-similarity coverage38.82-1.75
English-constrained coverage38.82-1.75
Script diversity40.03-0.12
COVER32.13+1.3

COVER improved held out residual access by 6.69 points over cosine selection and 4.71 over geographic coverage.

AuthorsOpenReview ↗

2. Larger language sets and new K values (8EM4, DEh3)

We fine tuned a new Qwen checkpoint on 10 languages (Arabic, English, French, Indonesian, Japanese, Korean, Nepali, Russian, Thai, Chinese) and used an expanded held out language set to include (Czech, German, Hindi, Italian, Dutch, Persian, Polish, Portuguese, Spanish, Turkish, Urdu, Vietnamese, Malay, Hebrew and Bengali) for a total of 25 languages. For K=4, COVER selected Arabic, English, Japanese and Nepali. Using the same fixed forget unlearning procedure as detailed in Section 2.

MethodLanguage SetSmall held out set ↓All held out languages ↓Retain drop ↓
COVERArabic/English/Japanese/Nepali0.1230.1790.151
FIXED COVEREnglish, Indonesian, Russian, Thai0.170.2560.16
Individual top-4Arabic, Indonesian, Japanese Nepali0.180.2800.156
Random (mean)See below0.2280.4270.106

Oracle normalized residual access

The small held out set equals the held out languages the checkpoint was trained on, but not unlearned on.

All held out languages are the small held out set plus the addition 15 listed above.

Random equals the average of:
French/Indonesian/Japanese/Chinese; French/Korean/Russian/Chinese; Arabic/Indonesian/Korean/Russian; English/French/Korean/Thai
Individual top-4 is the equivalent definition to individual Top-3 but in the K=4 setting.

We have added this experiment in Appendix D.3, Table 25, and summarize it in Section 4.3, in the paragraph.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.