Pith. sign in

REVIEW 4 major objections 6 minor 26 references

Mitigating Negative Interference in Multilingual Sequential Knowledge Editing through Null-Space Constraints

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A null-space projection of each language's parameter update keeps sequentially edited multilingual knowledge from interfering, preserving both edit accuracy and cross-lingual performance.

desk verdict LangEdit is a solid empirical extension of AlphaEdit to multilingual sequential editing, but the claimed mathematical guarantee is not established due to a projection dimension error and unaddressed feature drift. read the letter →

arxiv 2506.10800 v1 pith:5L3QEELA submitted 2025-06-12 cs.CL

Wei Sun , Tingyu Qu , Mingxiao Li , Jesse Davis , Marie-Francine Moens This is my paper · ORCID
classification cs.CL
keywords multilingualknowledgeeditingsequentialnull-spaceprojectionnegativeinterferencelargelanguagemodelscross-lingualgeneralizationfeaturecovariance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Sequential knowledge editing in a multilingual LLM normally lets a new edit in one language overwrite or degrade facts injected earlier in another language. LangEdit is built to stop that: each time it edits, it projects the parameter adjustment for the new language onto the null space of the feature covariance accumulated from all earlier language edits. The paper claims this makes updates independent by construction and shows, across three model families and six languages, that it beats adapted monolingual editing baselines on both editing accuracy and cross-lingual task performance. If right, a single multilingual model can be kept current across languages without per-language copies or retraining.

What carries the argument

The central object is the null-space projection matrix $P_{t-1}=U'_{t-1}(U'_{t-1})^\top$, built from the eigenvectors of the running uncentered feature covariance $\bar K_{t-1}$ whose eigenvalues are zero. It carries the argument by constraining each new weight update $\Delta W_t$ so that previously stored key-value associations are preserved exactly, while the current language's facts are learned in the orthogonal directions where earlier updates have no effect.

What would settle it

After each sequential edit, recompute the keys of all previously edited facts using the current model and check whether $\|(W_{t-1}+\Delta W_t)\bar K_{t-1}-\bar V_{t-1}\|$ stays near zero; if the residual grows with the number of edits, the claimed update-independence guarantee fails in practice.

Watch

Extended reading notes

Core claim

At each step $t$, LangEdit computes keys $K_t$ and values $V_t$ for the new language's facts, maintains a running uncentered covariance $\bar K_{t-1}$ of all previous keys, and projects the candidate weight update onto $P_{t-1}=U'_{t-1}(U'_{t-1})^\top$, the span of zero-eigenvalue eigenvectors of $\bar K_{t-1}$. The resulting closed-form update $\Delta W_t = R_t K_t^\top P_{t-1}(K_t K_t^\top P_{t-1}+I)^{-1}$ is constructed so that $(\Delta W_t+W_{t-1})\bar K_{t-1}=\bar V_{t-1}$, i.e., the edited model still maps all prior keys to their prior values. The paper's claim is that this mathematically guarantees update independence and, because the projection is rebuilt at every step from the growing covariance, it remains valid across a whole sequence of languages.

Load-bearing premise

The method assumes that the model's internal keys for old facts stay fixed after later edits change the model, so a null-space constraint built from earlier keys still describes old knowledge.

Editorial extensions

If this is right

  • A single multilingual LLM can absorb sequential edits in different languages without needing per-language copies or retraining, as long as a running feature covariance is maintained.
  • LangEdit's editing accuracy (Efficacy) exceeds adapted monolingual baselines by up to 2.20 percentage points on mzsre, and its cross-lingual F1 by up to 5.65 points.
  • The order in which languages are edited has little effect: shuffling the language order gives comparable results.
  • Facts injected in one language are answerable in other languages; the authors show cross-lingual transfer on queries never seen in the edits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: If key representations drift as edits accumulate, the stored covariance becomes stale; a direct test is to recompute old keys with the final model and measure the residual $\|W_T\bar K_{t-1}-\bar V_{t-1}\|$ per edit step.
  • Editorial extension: The same running-covariance projection could apply to other sequential update settings, such as continual instruction tuning or single-language fact streams, whenever each batch of updates has identifiable feature keys.
  • Editorial extension: The authors report that Spanish and French edits show larger gains than English edits, which suggests the method's headroom depends on how much pretraining data a language already has; one could test this by comparing low-resource and high-resource languages on the same fact set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces LangEdit, a method for multilingual sequential knowledge editing that constrains each parameter update to lie in an approximate null space of the feature covariance of previously edited knowledge. The authors define a benchmark built from Bi-ZsRE and M-ZsRE, evaluate on three LLMs (Llama3-8B, Qwen2.5-7B, GPT-J-6B) and four XTREME tasks, and report consistent small improvements over adapted baselines, with significance assessed by paired t-tests. The central methodological claim is that the null-space projection mathematically guarantees update independence across languages; the paper also reports ablations, per-language analyses, computational cost, and an additional MLaKE experiment.

Significance. If the mathematical guarantee were established, LangEdit would be a useful contribution: it addresses a real and under-studied problem, provides a reproducible benchmark for multilingual sequential editing, and shows consistent gains across architectures, languages, and downstream tasks. The paper has concrete strengths: the code is released, the evaluation is broad and includes ablation variants, per-language radar charts, and drift-relevant checks such as cross-lingual transfer experiments. However, the formal derivation of the null-space projection contains load-bearing dimension and equivalence errors, and the claimed exact guarantee is not supported as written. The empirical results may still be valid, but the contribution Section's 'provably' claim and the Abstract's 'mathematically guarantees' claim need a corrected derivation and an explicit treatment of representation drift.

major comments (4)
  1. [Section 3, Eqs. (7)-(9)] Under the definitions in Section 2.1, W and ΔW are d1×d0 matrices, while the projector P_{t-1} constructed from the SVD of the d0×d0 covariance \bar K_{t-1} is d0×d0. The expression P_{t-1}\tilde ΔW_t in Eqs. (8) and (9) is therefore undefined whenever d0≠d1, which is the case for all three evaluated models (e.g., in Llama3-8B the MLP has d0=14336 and d1=4096). The null-space constraint should be applied as \tilde ΔW_t P_{t-1}, or equivalently by projecting rows/columns of ΔW consistently. Consequently, the closed-form solution in Eq. (10) cannot be derived from Eq. (9); Eq. (10) multiplies P on the right in the first factor, which corresponds to a different optimization. This is a load-bearing error because the claimed 'provably' language-specific update rests on this derivation.
  2. [Section 2.2, Eq. (3)] Replacing the concatenated key and value matrices with their uncentered covariance matrices does not preserve Eq. (2). Eq. (2), (ΔW_t+W_{t-1})\bar K_{t-1}=\bar V_{t-1}, has left-hand side d1×d0 and right-hand side d1×n_{t-1}; Eq. (3), (ΔW_t+W_{t-1})\bar K_{t-1}=\bar V_{t-1} with covariance matrices, has left-hand side d1×d0 and right-hand side d1×d1. Moreover, (ΔW+W)\bar K=\bar V does not imply (ΔW+W)(\bar K^T\bar K)=\bar V^T\bar V; only the null-space condition ΔW x=0 for x in null(\bar K) follows from preservation of the old mapping. The paper should state the null-space constraint directly rather than presenting Eq. (3) as an equivalent reformulation.
  3. [Section 3, Eq. (6) and Algorithm 1, line 8] The instruction to keep eigenvectors of \bar K_{t-1} corresponding to zero eigenvalues is not operational. An empirical uncentered covariance matrix is generically full rank, so no exact zero eigenvalues exist; the method therefore requires an explicit tolerance or rank-selection rule. Without such a rule, the 'approximate null space' and the projector P_{t-1} are not uniquely defined, and the numerical behavior of the method depends on an undocumented threshold. The paper should specify the threshold, report its value, and include a sensitivity analysis over reasonable choices.
  4. [Section 3, Eq. (4) and Abstract] The 'mathematically guarantees update independence' claim is undermined by feature drift. The projector P_{t-1} is built from keys extracted under the parameter state at edit time t-1; when edits are applied to multiple critical layers, as in Section 4.1.3, later changes to W_out at lower layers alter hidden states and hence change the keys at higher layers for the same facts. The stored covariance \bar K_{t-1} is then stale with respect to the final model, so ΔW_t\bar K_{t-1}=0 does not imply that the final model preserves old facts. At a single edited layer the input keys are independent of W_out at that layer, so the issue arises mainly through multi-layer cascades, but the paper neither states this caveat nor bounds the drift. A concrete verification would be to re-extract old keys from the model after each edit and report the residual norm ||(W_t-W_{t-1})K_old^{(after)}||. Without such an analysis, the exact guarantee is unsupported; at best LangEdit provides a linearized protection whose accuracy depends on edit magnitudes.
minor comments (6)
  1. [Section 2.2, notation] The notation \bar K_{t-1}=[K_1;\cdots;K_{t-1}] suggests vertical stacking, which is inconsistent with K_t∈R^{d0×n_t}; please specify the concatenation direction explicitly in both Eq. (2) and the covariance updates.
  2. [Figure 3 caption] The caption reads 'We use , and to represent LangEdit, AlphaEdit and RECT' but the marker symbols are missing from the text; please insert the actual symbols.
  3. [Section 4.4, Figure 4] The sign convention for Δ is confusing: the text reports 'performance degradation' with positive Δ values for AlphaEdit (multi) and later reports positive values when LangEdit surpasses AlphaEdit (mono). Please state explicitly whether Δ is always defined as the performance gap in a fixed direction and keep the sign convention consistent.
  4. [Appendix A.2, Table 5] The text says the shuffled-language-order experiment was repeated three times, but Table 5 reports a single row without standard deviation; please report the mean and standard deviation over the three shuffles or clarify that only one representative shuffle is shown.
  5. [Appendix A.6] The MLaKE experiment uses 100 manually annotated subjects per language; this important limitation should be stated in the main text where the MLaKE results are mentioned, not only in the appendix.
  6. [Throughout] There are several typographical and punctuation errors, such as duplicated '): :' in Eq. (1) and missing symbols in the Figure 3 caption; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the null-space constraint is an explicit construction and all quantitative claims are evaluated on external benchmarks.

full rationale

LangEdit is a constrained optimization method: it solves Eqs. (5)-(10) with an explicit projection P_{t-1} onto the null space of the running covariance of previously edited keys. The identity (P_{t-1}\Delta W_t + W_{t-1})\bar K_{t-1} = \bar V_{t-1} is a consequence of the projection construction, not a fitted prediction. The reported empirical claims (Efficacy, Generality, Specificity, XTREME F1) are measured on external datasets that do not enter the projection or the optimization objective, so no evaluation target is used to fit the method. The null-space machinery is explicitly borrowed from Wang et al. (2021) and AlphaEdit (Fang et al., 2025), which are external to the present authors, so there is no load-bearing self-citation chain. The only substantive concern is that the guarantee is stated for keys extracted at edit time and is not re-verified after later edits change hidden states; this is an unanalyzed correctness/feature-drift issue, not circularity of the derivation. Under the provided rubric, no step reduces to its own input, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central method rests on the linear-memory view of MLPs, the exactness of old key-value constraints, and the stability of extracted keys under subsequent weight updates; none of these is independently verified in this paper. The numerical null-space threshold and hyperparameters are additional free choices. No new physical or architectural entities are introduced.

free parameters (5)
  • SVD zero-eigenvalue threshold = not reported
    Section 3 Eq. 6-7 keeps eigenvectors whose eigenvalues are exactly zero; in floating point no eigenvalue is exactly zero, so a threshold must be chosen but is not specified. This changes the projection subspace and therefore the update.
  • Number of optimization steps for key/value extraction = 25
    Section 4.1.3: 25 steps at learning rate 0.5 (GPT-J) or 0.1 (Llama3, Qwen2.5) to compute hidden representations. Copied from AlphaEdit/MEMIT but still a hand choice.
  • Learning rates for key extraction = 0.5 (GPT-J), 0.1 (Llama3, Qwen2.5)
    Section 4.1.3 hyperparameters for the optimization used to obtain keys and values for editing.
  • Critical edited layers = GPT-J [3,4,5,6,7,8]; Llama3 and Qwen2.5 [4,5,6,7,8]
    Selected via causal tracing (Appendix A.1); determines which MLP weight matrices are edited and is a modeling choice that affects all results.
  • Pre-edit key corpus size K0 = 100,000 random Wikipedia triplets
    Section 3, after Eq. 10; used to represent old knowledge. Its size and sampling affect the initial null space and hence all subsequent updates.
assumptions (4)
  • domain assumption MLP layers are linear associative memories and factual knowledge is stored as (subject, relation, object) in W_out
    Section 2.1 invokes Geva et al. 2021 and Meng et al. 2022; the choice to edit only W_out depends on this view of transformer knowledge storage.
  • standard math The null space of the uncentered covariance matrix equals the null space of the stacked keys
    Section 2.2 Eq. 3. True for a vector space if the covariance is computed with the correct orientation; the paper's notation is ambiguous about left versus right null space.
  • ad hoc to paper Old knowledge can be summarized by K0 from 100,000 random Wikipedia triplets, and W_{t-1} Kbar_{t-1} = Vbar_{t-1} holds exactly
    Section 3 Eq. 8 relies on previous weights satisfying the old key-value constraints exactly; approximate null space and feature drift make this an idealization.
  • domain assumption Keys extracted under earlier parameter states remain valid for later projection steps
    Section 3 Eq. 4 accumulates covariances without updating them after W changes; the paper does not analyze key drift, so the null-space constraint may not hold for the final model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Negative Interference in Multilingual Sequential Knowledge Editing through Null-Space Constraints." pith.science (2026). https://pith.science/paper/5L3QEELA

@misc{pith2026250610800,
  author       = {Pith},
  title        = {Pith review of: Mitigating Negative Interference in Multilingual Sequential Knowledge Editing through Null-Space Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5L3QEELA}},
  note         = {Machine review of arXiv:2506.10800}
}
read the original abstract

Efficiently updating multilingual knowledge in large language models (LLMs), while preserving consistent factual representations across languages, remains a long-standing and unresolved challenge. While deploying separate editing systems for each language might seem viable, this approach incurs substantial costs due to the need to manage multiple models. A more efficient solution involves integrating knowledge updates across all languages into a unified model. However, performing sequential edits across languages often leads to destructive parameter interference, significantly degrading multilingual generalization and the accuracy of injected knowledge. To address this challenge, we propose LangEdit, a novel null-space constrained framework designed to precisely isolate language-specific knowledge updates. The core innovation of LangEdit lies in its ability to project parameter updates for each language onto the orthogonal complement of previous updated subspaces. This approach mathematically guarantees update independence while preserving multilingual generalization capabilities. We conduct a comprehensive evaluation across three model architectures, six languages, and four downstream tasks, demonstrating that LangEdit effectively mitigates parameter interference and outperforms existing state-of-the-art editing methods. Our results highlight its potential for enabling efficient and accurate multilingual knowledge updates in LLMs. The code is available at https://github.com/VRCMF/LangEdit.git.

Figures

Figures reproduced from arXiv: 2506.10800 by the authors.

Figure 1
Figure 1. Illustration of Monolingual (a) and Multilin [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. F1 Scores of the edited Llama3-8B on XTREME benchmark evaluating multilingual generalization. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Radar chart of editing performance and multilingual downstream task performance when editing Llama3- [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance comparison for all languages between multilingual knowledge editing model and the state [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Average Indirect Effects of all model components (Llama3-8B) of 100 factual knowledge in six languages. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 13 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Lynn Chua, Badih Ghazi, Yangsibo Huang, Pritish Kamath, Ravi Kumar, Pasin Manurangsi, Amer Sinha, Chulin Xie, and Chiyuan Zhang. 2024. Crosslingual capabilities and knowledge barriers in multilingual large language models. arXiv preprint arXiv:2406.16135

    arXiv 2024

  4. [4]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. https://doi.org/10.18653/v1/2020.acl-main.747 Unsupervised cross-lingual representation learning at scale . In Proceedings of the 58th Annual Meeting of the Association for Comp...

    doi:10.18653/v1/2020.acl-main.747 2020

  5. [5]

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Xiang Wang, Xiangnan He, and Tat-seng Chua. 2025. Alphaedit: Null-space constrained knowledge editing for language models. International Conference on Learning Representations 2025

    work page 2025

  6. [6]

    Patrick Fernandes, Behrooz Ghorbani, Xavier Garcia, Markus Freitag, and Orhan Firat. 2023. Scaling laws for multilingual neural machine translation. In International Conference on Machine Learning, pages 10053--10071. PMLR

    work page 2023

  7. [7]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.446 Transformer feed-forward layers are key-value memories . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484--5495, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics

    doi:10.18653/v1/2021.emnlp-main.446 2021

  8. [8]

    Werner H Greub. 2012. Linear algebra, volume 23. Springer Science & Business Media

    work page 2012

Show all 26 references
  1. [9]

    Jia-Chen Gu, Hao-Xiang Xu, Jun-Yu Ma, Pan Lu, Zhen-Hua Ling, Kai-Wei Chang, and Nanyun Peng. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.934 Model editing harms general abilities of large language models: Regularization to the rescue . In Proceedings of the 2024 Conferen...

    2024 doi

  2. [10]

    Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. Smith. 2018. https://doi.org/10.18653/v1/N18-2017 Annotation artifacts in natural language inference data . In Proceedings of the 2018 Conference of the North A merican Chapter of the As...

    2018 doi

  3. [11]

    Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan Firat, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation. In International Conference on Machine Learning, pages 4411--4421. PMLR

    2020

  4. [12]

    Patrick Lewis, Barlas Oguz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk. 2020. https://doi.org/10.18653/v1/2020.acl-main.653 MLQA : Evaluating cross-lingual extractive question answering . In Proceedings of the 58th Annual Meeting of the Association for Computational Lin...

    2020 doi

  5. [13]

    Jun-Yu Ma, Hong Wang, Hao-Xiang Xu, Zhen-Hua Ling, and Jia-Chen Gu. 2025. Perturbation-restrained sequential model editing. International Conference on Learning Representations 2025

    2025

  6. [14]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359--17372

    2022

  7. [15]

    Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. 2023. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations

    2023

  8. [16]

    Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. https://doi.org/10.18653/v1/P17-1178 Cross-lingual name tagging and linking for 282 languages . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (...

    2017 doi

  9. [17]

    Aryan Singhal, Thomas Law, Coby Kassner, Ayushman Gupta, Evan Duan, Aviral Damle, and Ryan Luo Li. 2024. https://doi.org/10.18653/v1/2024.nlp4pi-1.2 Multilingual fact-checking using LLM s . In Proceedings of the Third Workshop on NLP for Positive Impact, pages 13--31, Miami, F...

    2024 doi

  10. [18]

    Russell A Smith. 1967. The condition numbers of the matrix eigenvalue problem. Numerische Mathematik, 10:232--240

    1967

  11. [19]

    Jiaan Wang, Yunlong Liang, Zengkui Sun, Yuxuan Cao, Jiarong Xu, and Fandong Meng. 2024 a . https://doi.org/10.18653/v1/2024.acl-long.627 Cross-lingual knowledge editing in large language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Li...

    2024 doi

  12. [20]

    Shipeng Wang, Xiaorong Li, Jian Sun, and Zongben Xu. 2021. Training networks in null space of feature covariance for continual learning. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 184--193

    2021

  13. [21]

    Weixuan Wang, Barry Haddow, and Alexandra Birch. 2024 b . https://doi.org/10.18653/v1/2024.acl-long.21 Retrieval-augmented multilingual knowledge editing . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages...

    2024 doi

  14. [22]

    Yabing Wang, Fan Wang, Jianfeng Dong, and Hao Luo. 2024 c . Cl2cm: Improving cross-lingual cross-modal retrieval via cross-lingual knowledge transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5651--5659

    2024

  15. [23]

    Zihao Wei, Jingcheng Deng, Liang Pang, Hanxing Ding, Huawei Shen, and Xueqi Cheng. 2025. https://aclanthology.org/2025.coling-main.301/ ML a KE : Multilingual knowledge editing benchmark for large language models . In Proceedings of the 31st International Conference on Computa...

    2025

  16. [24]

    Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. https://doi.org/10.18653/v1/D19-1382 PAWS - X : A cross-lingual adversarial dataset for paraphrase identification . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the ...

    2019 doi

  17. [25]

    Fuwei Zhang, Zhao Zhang, Xiang Ao, Dehong Gao, Fuzhen Zhuang, Yi Wei, and Qing He. 2022. Mind the gap: Cross-lingual information retrieval with hierarchical knowledge enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 4345--4353

    2022

  18. [26]

    Xue Zhang, Yunlong Liang, Fandong Meng, Songming Zhang, Yufeng Chen, Jinan Xu, and Jie Zhou. 2025. https://aclanthology.org/2025.coling-main.385/ Multilingual knowledge editing with language-agnostic factual neurons . In Proceedings of the 31st International Conference on Comp...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.