pith. sign in

arxiv: 2606.03291 · v1 · pith:Y7DWSQHKnew · submitted 2026-06-02 · 💻 cs.CL

Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility

Chaoyi Xiang , Olga Ohrimenko , Benjamin I. P. Rubinstein , Lea Frermann This is my paper

Pith reviewed 2026-06-28 10:15 UTC · model grok-4.3

classification 💻 cs.CL
keywords multilingual unlearninglarge language modelsTOFU benchmarklayer-wise analysisinference-time steeringcross-lingual transferknowledge suppressionreversibility
0
0 comments X

The pith

Unlearning in LLMs suppresses knowledge superficially in later layers rather than erasing it from early shared representations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper extends the TOFU benchmark to five languages to examine how unlearning targeted facts transfers across languages in LLMs. Experiments with different language combinations for fine-tuning, unlearning, and querying show that transfer strength depends on script and language family overlap, with the unlearning language itself predicting the strongest transfer targets. Layer-wise probing finds that early layers keep the shared cross-lingual space intact while later decoding layers perform the suppression. This structure implies unlearning is not erasure but reversible surface-level change. A single inference-time steering direction recovers 50 percent of the knowledge in Qwen and 90 percent in Gemma across languages.

Core claim

Unlearning leaves the shared cross-lingual latent space largely intact in early layers and operates primarily in later decoding layers. This indicates that unlearning induces superficial suppression instead of true erasure of knowledge. Consequently, a single inference-time steering direction can reverse much of this suppression, recovering 50% of the unlearned knowledge in Qwen and 90% in Gemma across languages.

What carries the argument

Layer-wise analysis showing unlearning affects later decoding layers while preserving early cross-lingual representations, which enables reversal by a single steering direction at inference time.

If this is right

  • Unlearning transfer is strongest between languages that share scripts and families.
  • The language chosen for unlearning predicts which query languages will show the largest transfer effects.
  • Early layers preserve the shared latent space even after unlearning in one language.
  • A single steering vector at inference time can recover substantial portions of suppressed knowledge without retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Existing unlearning methods may fail to provide durable privacy guarantees once models are queried in multiple languages.
  • Steering vectors offer a lightweight way to toggle access to previously unlearned content after deployment.
  • The observed layer separation could apply to other suppression techniques used for safety or copyright control.
  • Measuring transfer on languages with greater typological distance would test whether the reported patterns hold beyond the five-language set.

Load-bearing premise

The TOFU facts and their translations remain semantically equivalent and equally memorable across the five languages.

What would settle it

Finding that the steering direction recovers none of the unlearned facts when tested on held-out models or language pairs, or that early layers exhibit large representation shifts after unlearning, would falsify the superficial-suppression account.

Figures

Figures reproduced from arXiv: 2606.03291 by Benjamin I. P. Rubinstein, Chaoyi Xiang, Lea Frermann, Olga Ohrimenko.

Figure 1
Figure 1. Figure 1: Layer-wise cosine similarity of final-token hidden states h (l) m (x). Top row: cross-language comparison between embeddings of an English question and its translation (Ch, De, Ru, Tu) in the same model (Qwen2.5-7B variants; see panel titles). FT = fine-tuning, UN = unlearning and Cross-lingual prompting means that for non-English questions, the model is instructed to answer in English. Bottom row: cross-m… view at source ↗
Figure 2
Figure 2. Figure 2: PCA separation (Qwen2.5-7B) across layers for English (top) and Chinese (bottom) questions. Numbers indicate question indices, with identical numbers referring to the same questions across different testing model variants. PCA analysis to German, Russian, and Turkish as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of layer-wise steering vector injection on NLI score for the forget set. Left: NLI score change when injecting a scaled and normalized steering vector extracted from the retain set, showing substantial recovery (positive score gain). Right: NLI score changes when injecting random Gaussian directions with matched norm and scale. The much weaker recovery indicates that the extracted vectors capture a … view at source ↗
Figure 4
Figure 4. Figure 4: Layer-wise cosine similarity between the hidden state h (l) m (xℓ) of the final input token for Chinese, English, German, Russian, and Turkish questions, across three variants of Qwen2.5-7B (see panel titles for settings). FT denotes fine-tuning, and UN denotes unlearning. Cross-lingual prompting means that for non-English questions, the model is instructed to answer in English. Bottom row: control compari… view at source ↗
Figure 5
Figure 5. Figure 5: We compare the cosine similarity of hidden representations before and after unlearning within the same language for which the model was finetuned. The results show that the early and mid layers are minimally affected, while the majority of the similarity drop occurs in the later layers. extending the analysis to the Base model (FT vs Base and UN vs Base) confirms that the unlearned model does not simply re… view at source ↗
Figure 6
Figure 6. Figure 6: PCA separation across layers for German (top), Russian (middle) and Turkish (bottom) questions. Numbers indicate question indices, with identical numbers referring to the same questions across different testing model variants. K. Additional Steering Vector Experiments K.1. Chinese as the Source Language In the main text, we construct steering vectors using English questions as the source language. To test … view at source ↗
Figure 7
Figure 7. Figure 7: The left heatmap capture the impact of unlearning for global geometry, while the right heatmap capture per-sample effect of unlearning. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 Layer English Chinese German Russian Turkish Language Injecting Extracted Steering Vector 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 … view at source ↗
Figure 8
Figure 8. Figure 8: Gemma2-9B steering injection results. Effect of layer-wise steering on NLI score for the forget set. Each heatmap shows the change in score when injecting a normalized steering direction with scale α, where α is selected separately for each language. The left heatmap uses extracted steering vectors, which recover almost all of the forgotten knowledge when applied in middle and late layers. The right heatma… view at source ↗
read the original abstract

Large language models (LLMs) can memorize sensitive facts, motivating unlearning methods that remove targeted knowledge without costly retraining. However, unlearning research remains heavily English-centric. We study multilingual unlearning by extending the TOFU benchmark to five languages, and fine-tune, unlearn, and query our models with different permutations of languages. We find that unlearning transfer, the ability of an unlearned model to "forget" facts in languages other than the unlearning language, is highly variable: e.g., it is strongest between languages sharing scripts and families, and we show that the unlearning language predicts which query languages are most likely to yield the strongest transfer. Layer-wise analysis reveals that unlearning leaves the shared cross-lingual latent space largely intact in early layers, instead operating primarily in later decoding layers. This suggests that unlearning does not truly erase knowledge, but rather induces superficial suppression. Exploiting this structure, a single inference-time steering direction reverses much of this suppression across languages, recovering 50% (Qwen) and 90% (Gemma) of the unlearned knowledge.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper extends the TOFU benchmark to five languages and examines multilingual unlearning in LLMs via fine-tuning and unlearning experiments across language permutations. It reports highly variable unlearning transfer (strongest between script- and family-sharing languages), shows via layer-wise analysis that unlearning primarily affects later decoding layers while leaving early cross-lingual representations intact, and demonstrates that a single inference-time steering vector recovers 50% (Qwen) to 90% (Gemma) of unlearned knowledge across languages. The central interpretation is that unlearning induces superficial suppression rather than true erasure.

Significance. If the empirical patterns hold after addressing benchmark controls, the work would usefully document the limits of unlearning in multilingual settings and identify a practical reversal mechanism via steering. The layer-wise findings offer a concrete mechanistic hypothesis that could guide future unlearning research, though the absence of statistical details and translation validation currently limits the strength of the cross-lingual claims.

major comments (2)
  1. [Abstract] Abstract: The central claims of variable transfer strength and steering-based recovery (50% Qwen, 90% Gemma) rest on the assumption that the extended TOFU facts remain semantically equivalent, equally difficult, and culturally salient across the five languages. No validation of translations, inter-annotator agreement, or controls for translation artifacts are described, which directly risks confounding the reported language-family effects and recovery percentages with input differences rather than model-internal suppression.
  2. [Abstract] Abstract and experimental description: The soundness assessment notes the absence of statistical significance testing, exact layer definitions, and translation-quality controls. These omissions are load-bearing because the layer-wise analysis and cross-lingual transfer conclusions are presented as clear patterns without reported error bars, p-values, or ablation on translation fidelity.
minor comments (1)
  1. [Abstract] Abstract: The phrase 'we show that the unlearning language predicts which query languages are most likely to yield the strongest transfer' would benefit from a precise definition of the prediction metric and the supporting figure or table.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the need for stronger controls on translation fidelity and statistical reporting. We address each concern below and will revise the manuscript accordingly to improve clarity and robustness.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claims of variable transfer strength and steering-based recovery (50% Qwen, 90% Gemma) rest on the assumption that the extended TOFU facts remain semantically equivalent, equally difficult, and culturally salient across the five languages. No validation of translations, inter-annotator agreement, or controls for translation artifacts are described, which directly risks confounding the reported language-family effects and recovery percentages with input differences rather than model-internal suppression.

    Authors: We agree that explicit validation of the translated TOFU facts would strengthen the cross-lingual claims. The current manuscript does not report inter-annotator agreement or systematic controls for translation artifacts. In revision we will add a dedicated appendix describing the translation pipeline (professional translators plus back-translation checks) and any available quality metrics. We note, however, that the observed transfer patterns align with independent linguistic distances (script and family overlap) rather than arbitrary translation quality, which provides some indirect support that the effects are model-internal; this argument will be made more explicit. revision: yes

  2. Referee: [Abstract] Abstract and experimental description: The soundness assessment notes the absence of statistical significance testing, exact layer definitions, and translation-quality controls. These omissions are load-bearing because the layer-wise analysis and cross-lingual transfer conclusions are presented as clear patterns without reported error bars, p-values, or ablation on translation fidelity.

    Authors: We accept that the absence of error bars, p-values, and precise layer indexing weakens the presentation. In the revised manuscript we will (i) report standard errors and two-sided t-tests for all transfer and recovery percentages, (ii) define layers explicitly (early: layers 0–8, middle: 9–16, late: 17–23 for the 24-layer models) with a figure showing the exact indices used, and (iii) include an ablation that recomputes key metrics on a subset of facts whose translations were independently verified by a second annotator. These additions directly address the load-bearing concerns. revision: yes

Circularity Check

0 steps flagged

Empirical study with no derivations or self-referential reductions

full rationale

The paper reports experimental measurements of unlearning transfer, layer-wise activation differences, and steering-vector recovery on an extended TOFU benchmark. All reported percentages (e.g., 50% Qwen, 90% Gemma recovery) and qualitative findings are direct outcomes of fine-tuning, evaluation, and probing; none are obtained by fitting a parameter to a subset and relabeling it a prediction, nor by self-citation of a uniqueness theorem, nor by any equation that reduces to its own inputs by construction. The work contains no mathematical derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are introduced; the work is purely empirical measurement on existing model architectures and a translated benchmark.

pith-pipeline@v0.9.1-grok · 5733 in / 1111 out tokens · 16912 ms · 2026-06-28T10:15:13.244733+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 16 canonical work pages · 1 internal anchor

  1. [1]

    acl-short.103/

    URL https://aclanthology.org/2023. acl-short.103/. Anna, B., Savchenko, A., Panchenko, A., and Tutubalina, E. Anatomy of unlearning: The dual impact of fact salience and model fine-tuning, 2026. URL https: //arxiv.org/abs/2602.19612. Blaschke, V ., Fedzechkina, M., and Ter Hoeve, M. Ana- lyzing the effect of linguistic similarity on cross-lingual transfer...

    Pith/arXiv arXiv 2023

  2. [3]

    findings-acl.454/

    URL https://aclanthology.org/2025. findings-acl.454/. Cao, Y . and Yang, J. Towards making systems forget with machine unlearning. In2015 IEEE Symposium on Security and Privacy, pp. 463–480, 2015. doi: 10.1109/SP.2015.35. Carlini, N., Ippolito, D., Jagielski, M., Lee, K., Tramer, F., and Zhang, C. Quantifying memorization across neural language models. In...

    work page doi:10.1109/sp.2015.35 2025

  3. [4]

    ISBN 979-8-89176-256-5

    Association for Computational Linguistics. ISBN 979-8-89176-256-5. doi: 10.18653/v1/2025.findings-acl

    work page doi:10.18653/v1/2025.findings-acl 2025

  4. [5]

    findings-acl.310/

    URL https://aclanthology.org/2025. findings-acl.310/. Chang, T. A., Arnett, C., Tu, Z., and Bergen, B. K. When is multilinguality a curse? language modeling for 250 high- and low-resource languages. In Al- Onaizan, Y ., Bansal, M., and Chen, Y .-N. (eds.),Pro- ceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 4074–40...

    work page doi:10.18653/v1/2024.emnlp-main 2025

  5. [6]

    emnlp-main.236/

    URL https://aclanthology.org/2024. emnlp-main.236/. Chen, Y . and Eger, S. MENLI: Robust evaluation metrics from natural language inference.Transactions of the Association for Computational Linguistics, 11:804–825,

    2024

  6. [7]

    URL https:// aclanthology.org/2023.tacl-1.47/

    doi: 10.1162/tacl_a_00576. URL https:// aclanthology.org/2023.tacl-1.47/. Cheng, A., Huang, W., and Wang, Y . A fully proba- bilistic perspective on large language model unlearn- ing: Evaluation and optimization. In Christodoulopou- los, C., Chakraborty, T., Rose, C., and Peng, V . (eds.),Proceedings of the 2025 Conference on Em- pirical Methods in Natura...

    work page doi:10.1162/tacl_a_00576 2023

  7. [8]

    emnlp-main.452/

    URL https://aclanthology.org/2025. emnlp-main.452/. Choi, M., Min, K., and Choo, J. Cross-lingual unlearning of selective knowledge in multilingual language models. In Al-Onaizan, Y ., Bansal, M., and Chen, Y .-N. (eds.), 10 Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility Findings of the Association for Computational Linguis- tics: ...

    work page doi:10.18653/v1/2024.findings-emnlp 2025

  8. [9]

    XNLI : Evaluating Cross-lingual Sentence Representations

    URL https://aclanthology.org/2024. findings-emnlp.630/. Conneau, A., Rinott, R., Lample, G., Williams, A., Bowman, S. R., Schwenk, H., and Stoyanov, V . XNLI: Evaluat- ing cross-lingual sentence representations. In Riloff, E., Chiang, D., Hockenmaier, J., and Tsujii, J. (eds.),Pro- ceedings of the 2018 Conference on Empirical Methods in Natural Language P...

    work page doi:10.18653/v1/d18-1269 2024

  9. [10]

    Dorna, V ., Mekala, A

    URL https://openreview.net/forum? id=X39dK0SX9W. Dorna, V ., Mekala, A. R., Zhao, W., McCallum, A., Kolter, J. Z., Lipton, Z. C., and Maini, P. Openunlearning: Ac- celerating LLM unlearning via unified benchmarking of methods and metrics. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2026. URL...

    work page doi:10.18653/v1/2020.inlg-1 2026

  10. [11]

    How Contextual are Contextualized Word Representations? C omparing the Geometry of BERT , ELM o, and GPT -2 Embeddings

    URL https://aclanthology.org/2020. inlg-1.19/. Ethayarajh, K. How contextual are contextualized word rep- resentations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings. In Inui, K., Jiang, J., Ng, V ., and Wan, X. (eds.),Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Confer...

    work page doi:10.18653/v1/d19-1006 2020

  11. [12]

    Hu, S., Fu, Y ., Wu, S., and Smith, V

    URL https://proceedings.mlr.press/ v119/hu20b.html. Hu, S., Fu, Y ., Wu, S., and Smith, V . Unlearning or ob- fuscating? jogging the memory of unlearned LLMs via benign relearning. InThe Thirteenth International Confer- ence on Learning Representations, 2025. URL https: //openreview.net/forum?id=fMNRYBvcQN. Huang, J. Y ., Zhou, W., Wang, F., Morstatter, F...

    work page doi:10.1609/aaai.v39i22 2025

  12. [13]

    Nguyen, See-Kiong Ng, and Anh Tuan Luu

    URL https://aclanthology.org/2023. acl-long.805/. Jia, J., Zhang, Y ., Zhang, Y ., Liu, J., Runwal, B., Diffend- erfer, J., Kailkhura, B., and Liu, S. SOUL: Unlocking the power of second-order optimization for LLM unlearn- ing. In Al-Onaizan, Y ., Bansal, M., and Chen, Y .-N. (eds.),Proceedings of the 2024 Conference on Empiri- cal Methods in Natural Lang...

    work page doi:10.18653/v1/2024 2023

  13. [14]

    K, K., Wang, Z., Mayhew, S., and Roth, D

    URL https://openreview.net/forum? id=wOmtZ5FgMH. K, K., Wang, Z., Mayhew, S., and Roth, D. Cross-lingual ability of multilingual bert: An empirical study. In International Conference on Learning Representations,

  14. [15]

    URL https://openreview.net/forum? id=HJeT3yrtDr. Kang, E. and Kim, J. When language shapes thought: Cross-lingual transfer of factual knowledge in question answering. InProceedings of the 34th ACM Interna- tional Conference on Information and Knowledge Man- agement, CIKM ’25, pp. 4868–4873. ACM, November

  15. [16]

    URL http: //dx.doi.org/10.1145/3746252.3760807

    doi: 10.1145/3746252.3760807. URL http: //dx.doi.org/10.1145/3746252.3760807. Kornblith, S., Norouzi, M., Lee, H., and Hinton, G. Sim- ilarity of neural network representations revisited. In Chaudhuri, K. and Salakhutdinov, R. (eds.),Proceed- ings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learn- ing Resea...

    work page doi:10.1145/3746252.3760807 2019

  16. [18]

    emnlp-main.1109/

    URL https://aclanthology.org/2025. emnlp-main.1109/. 12 Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility Li, K., Patel, O., Viégas, F., Pfister, H., and Watten- berg, M. Inference-time intervention: Eliciting truth- ful answers from a language model. InThirty-seventh Conference on Neural Information Processing Systems,

    2025

  17. [19]

    Li, N., Pan, A., Gopal, A., Yue, S., Berrios, D., Gatti, A., Li, J

    URL https://openreview.net/forum? id=aLLuYpn83y. Li, N., Pan, A., Gopal, A., Yue, S., Berrios, D., Gatti, A., Li, J. D., Dombrowski, A.-K., Goel, S., Mukobi, G., Helm-Burger, N., Lababidi, R., Justen, L., Liu, A. B., Chen, M., Barrass, I., Zhang, O., Zhu, X., Tamirisa, R., Bharathi, B., Herbert-V oss, A., Breuer, C. B., Zou, A., Mazeika, M., Wang, Z., Osw...

  18. [20]

    URL https://proceedings.mlr.press/ v235/li24bc.html. Lim, Z. W., Aji, A. F., and Cohn, T. Language-specific la- tent process hinders cross-lingual performance.arXiv preprint arXiv:2505.13141, 2025. URL https:// arxiv.org/abs/2505.13141. Lin, C.-Y . ROUGE: A package for automatic evalua- tion of summaries. InText Summarization Branches Out, pp. 74–81, Barc...

    work page doi:10.18653/v1/p19-1301 2025

  19. [21]

    emnlp-main.516/

    URL https://aclanthology.org/2025. emnlp-main.516/. Lynch, A., Guo, P., Ewart, A., Casper, S., and Hadfield- Menell, D. Eight methods to evaluate robust unlearn- ing in llms, 2024. URL https://arxiv.org/abs/ 2402.16835. Maini, P., Feng, Z., Schwarzschild, A., Lipton, Z. C., and Kolter, J. Z. TOFU: A task of fictitious unlearning for LLMs. InFirst Conferen...

    arXiv 2025

  20. [22]

    Meng, K., Bau, D., Andonian, A

    URL https://openreview.net/forum? id=B41hNBoWLo. Meng, K., Bau, D., Andonian, A. J., and Belinkov, Y . Lo- cating and editing factual associations in GPT. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K. (eds.), Advances in Neural Information Processing Systems,

  21. [23]

    URL https://openreview.net/forum? id=-h6WAS6eE4. Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., Mann, B., Askell, A., Bai, Y ., Chen, A., Conerly, T., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Johnston, S., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Clark, J., Kaplan, J., McCandlish...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/ 2022

  22. [24]

    Lost in Multilinguality: Dissecting Cross-lingual Factual Inconsistency in Transformer Language Models

    doi: 10.18653/v1/2025.acl-long.253. URL https: //aclanthology.org/2025.acl-long.253/. Wang, Y ., Wei, J., Liu, C. Y ., Pang, J., Liu, Q., Shah, A., Bao, Y ., Liu, Y ., and Wei, W. LLM unlearning via loss adjustment with only forget data. InThe Thirteenth International Conference on Learning Representations, 2025b. URL https://openreview.net/forum? id=6ESR...

    work page doi:10.18653/v1/2025.acl-long.253 2025

  23. [25]

    Do Llamas Work in

    URL https://openreview.net/forum? id=BmEH70Wjcu. Wendler, C., Veselovsky, V ., Monea, G., and West, R. Do llamas work in English? on the latent language of multi- 14 Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility lingual transformers. In Ku, L.-W., Martins, A., and Sriku- mar, V . (eds.),Proceedings of the 62nd Annual Meeting of th...

    work page doi:10.18653/v1/2024.acl-long.820 2024

  24. [27]

    emnlp-main.265/

    URL https://aclanthology.org/2025. emnlp-main.265/. Yoon, Y ., Nam, J., Yun, H., Lee, J., Kim, D., and Ok, J. Few-shot unlearning by model inversion, 2023. URL https://arxiv.org/abs/2205.15567. Zhang, R., Lin, L., Bai, Y ., and Mei, S. Negative preference optimization: From catastrophic collapse to effective un- learning. InFirst Conference on Language Modeling,

    arXiv 2025

  25. [28]

    naacl-long.236/

    URL https://openreview.net/forum? id=MXLBXjQkmb. Zhao, Y ., Zhang, W., Chen, G., Kawaguchi, K., and Bing, L. How do large language models handle multilingual- ism? InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https: //openreview.net/forum?id=ctXYOoAgRy. Zhu, Y ., Liu, D., Lin, Z., Tong, W., Zhong, S., and Shao,...

    work page doi:10.18653/v1/2025.emnlp-main 2024

  26. [29]

    I cannot answer this question

    URL https://aclanthology.org/2025. emnlp-main.61/. 15 Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility A. Data Contamination Discussion TOFU was introduced in January 2024 (Maini et al., 2024) (arXiv submission: 11 Jan 2024), and its author profiles are synthetically generated, making pretraining contamination unlikely. Community dis...

    2025