Pith. sign in

REVIEW 3 major objections 5 minor 58 references

Mass-Editing Memory with Attention in Transformers: A cross-lingual exploration of knowledge

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Adding a small set of attention-head corrections on top of MEMIT editing improves every measured knowledge-editing metric, raises magnitude metrics by about ten percentage points, and transfers to languages and fact sets it was not…

desk verdict A cheap, sensible add-on to MEMIT with a genuinely portable recycled-head result, but the paper oversells the mechanistic claim without a random-head control. read the letter →

arxiv 2502.02173 v1 pith:2I655BOS submitted 2025-02-04 cs.CL cs.AI

Daniel Tamayo , Aitor Gonzalez-Agirre , Javier Hernando , Marta Villegas This is my paper · ORCID
classification cs.CLcs.AI
keywords knowledgeeditingmassattentionheadscross-lingualtransfermultilinguallanguagemodelsfactualassociationsMEMATCounterFact
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

The paper proposes MEMAT, which adds optimized corrections to a handful of attention heads after the feed-forward-based editor MEMIT has inserted new facts. Its claim is that these small corrections improve every measured outcome of knowledge editing—how often the new fact is produced, how well it generalizes to paraphrases, and how well unrelated facts are preserved—with gains of about ten percentage points on magnitude metrics, which gauge the model's confidence in the new answer. The method is shown to work when training and evaluation use different languages (English and Catalan) and when corrections are recycled into edits of entirely different facts, which the authors take as evidence that attention heads encode editing behavior at a cross-lingual level.

What carries the argument

The load-bearing object is a set of per-head additive value corrections $\omega^{\ell,h}$ inserted into the multi-query self-attention output before the output projection. The method selects the top-$K$ heads by per-head sigmoid classifiers trained on the last subject token's attention-head outputs after MEMIT editing, and optimizes the corrections with a loss that combines an edit objective with a KL-divergence term to preserve unrelated knowledge. This is what carries the argument: a small, portable set of attention parameters rather than further feed-forward rewriting.

What would settle it

Train the head classifiers and corrections on a language pair that does not share an alphabet or shared subject tokenization (for example, English and Arabic), or swap MEMIT for a different MLP-based editor as the base, and check whether the same 16 head corrections still beat the base editor on all six metrics. A clear failure in either test would refute the portability claim that carries the method.

Watch

Extended reading notes

Core claim

The paper's central discovery is that after MEMIT has inserted up to thousands of factual associations by modifying feed-forward layers in one language, the remaining gap in the model's confidence and generalization can be closed by optimizing additive corrections $\omega^{\ell,h}$ for the top $K=16$ attention heads. Those heads are chosen by training a sigmoid classifier per head on the last-subject-token representations of true versus false paraphrases in a second language, then selecting the heads whose predictions best match the edited facts. The corrections are optimized with a loss that maximizes the probability of the new object under randomized prompts while keeping a KL-divergence penalty that protects unrelated knowledge. In experiments with English and Catalan on a 7B decoder-only model, MEMAT outperforms MEMIT on all success and magnitude metrics, with the largest gains in paraphrase magnitude, and the corrections remain effective when applied to a different set of inserted facts or evaluated in the language not used for training.

Load-bearing premise

MEMAT assumes that the small set of attention components found useful after one batch of edits in one language is also the right set to correct for other facts and other languages; if that localization fails to carry over, the reported gains disappear.

Editorial extensions

If this is right

  • Editors can now improve edited-fact confidence without modifying more feed-forward weights: 16 attention-head corrections beat MEMIT on efficacy, paraphrase, and neighborhood metrics in both English and Catalan.
  • Knowledge editing can be performed in one language and take effect in another, at least when subjects tokenize similarly across languages, so multilingual updates could be issued once rather than repeated per language.
  • Head corrections are portable to new fact sets, so the expensive head-localization step does not have to be rerun for every batch of edits.
  • The gains hold as the number of inserted facts grows from roughly 100 to 10,000, although both MEMIT and MEMAT still degrade with scale.
  • Inserting facts bilingually with merged MEMIT matrices plus MEMAT corrections raises paraphrase metrics but lowers neighborhood metrics relative to monolingual insertion, pointing to a trade-off between cross-lingual understanding and specificity.

Reading between the lines

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

  • A direct test the paper does not run is whether the corrections can be combined with PMET or with a different base editor before the MEMIT step; since the selection procedure only needs an edited model, the recipe may generalize to other editors.
  • The strong cross-lingual transfer on English-Catalan suggests a follow-up on a non-alphabet-sharing pair; the paper's own limitations section says all conclusions come from these two similar languages.
  • The cost asymmetry (about 25 GPU minutes for head corrections versus 3 GPU hours for MEMIT) suggests a reusable correction bank could make repeated editing much cheaper, although the paper does not calculate end-to-end savings.
  • Because head corrections trained on one fact set sometimes beat freshly trained ones, the corrections may be learning a general 'how to edit' direction rather than the specific content of edited facts; probing the geometry of the correction vectors would test this.
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

3 major / 5 minor

Summary. The paper studies cross-lingual knowledge editing in a large language model (Aguila-7B) by first analyzing MEMIT in English and Catalan, finding that performance depends on subject-tokenization similarity. It then proposes MEMAT, which after a MEMIT edit locates attention heads using sigmoid classifiers on the last token of edited triplets (Section 5.2), optimizes additive per-head corrections on the K=16 highest-scoring heads under a loss that includes a KL term (Eq. 12), and applies these corrections during inference. The paper reports that MEMAT consistently outperforms MEMIT across success, magnitude, and accuracy metrics in Tables 1-3, including experiments where the head corrections are recycled to new fact sets and languages (L1-L2* cases).

Significance. If the empirical claims hold, the paper shows that a small set of additive attention-head corrections can improve the fidelity of mass-edited factual knowledge while remaining portable to new facts and to a language not used for the head optimization. The manuscript has two clear strengths: it ships code and data via GitHub, and the recycled-head experiments (Section 6, Tables 1-2, asterisked cases) provide a genuine external validation of head transfer rather than only fitting the training instances. The cross-lingual localization maps (Figures 2 and 9) also provide useful descriptive evidence. However, the significance is bounded by evaluation on a single model and two closely related languages, and by the absence of a control for whether the located heads themselves are responsible for the gains.

major comments (3)
  1. [Section 6(c), Tables 1-3] The central mechanistic claim that the sigmoid-classifier-located heads are causes of MEMAT's improvement is untested. All reported MEMAT variants optimize K=16 heads chosen by the Section 5.2 procedure, but there is no comparison in which K=16 random heads, or K=16 lowest-scoring heads, are optimized with the same loss (Eq. 12), the same regularization, and the same optimization budget. The recycled-head experiments (L1-L2*) show transfer, but a random-head adapter could transfer equally well. Without this control, the gains over MEMIT could be attributable to the added correction capacity rather than to the head-localization rationale, and the conclusion in Section 8 that attention heads encode information in a language-independent manner is not supported. Please add this control.
  2. [Appendix F and Section 6] The choice K=16 is selected by evaluating the same metrics that appear in the main tables. Appendix F (Figure 10) shows all six success and magnitude metrics for K in {8,16,32,48} on the same evaluation setup, and the text states that the optimal number is around K=16. This means the reported MEMAT-16 results are partly the product of selecting K on the evaluation set, so the claim of consistent improvement over MEMIT is optimistic. Please either pre-register K, select it on an independent validation split that is not used in the reported tables, or report the full K-sweep with selection costs clearly separated from the final evaluation.
  3. [Abstract and Table 1] The abstract's statement of 'significant improvements in all metrics' is stronger than the reported 95% confidence intervals support. In Table 1, for example, English ES for MEMAT-16 (CC) is 90.3 (1.1) versus 88.8 (0.7) for MEMIT (CAT), and English NS is 72.7 (1.2) versus 71.8 (0.8); these intervals overlap. The claim may hold for magnitude metrics, where the differences are larger, but it does not hold for all success metrics at the 95% level. Please either provide paired significance tests on the underlying samples or revise the wording to 'improves on all metrics' and reserve 'significant' for the metrics where the intervals do not overlap.
minor comments (5)
  1. [Eq. (11)] The notation 'K args max' is nonstandard; please rewrite to define an argmax over the K highest-scoring positions and clarify the role of the validation fraction beta in the computation.
  2. [Eq. (12) and Section 6(c)] The hyperparameters of the head-correction optimization, including lambda_omega, the number of random prompts R, the Adam batch size, and gradient-accumulation steps, are not reported. Please provide these values for reproducibility.
  3. [Appendix G] The accuracy metrics EA, PA, and NA are introduced only in Appendix G but are used to describe the main results; please define them in Section 4 or explicitly state that they are defined in the appendix.
  4. [Abstract and Section 6] The abstract claims a 'remarkable 10% increase in magnitude metrics', but the tables show much larger relative gains on some magnitude metrics (e.g., English PM in Table 1 increases from 22.2 to 27.8, about 25%). Please specify the baseline and metric used for the 10% figure.
  5. [Figure 3 caption] The note that the Eagle images were generated by GPT-4 is unrelated to the method and can be removed for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MEMAT's claims are supported by external portability experiments, not by equation-level reduction to its inputs.

full rationale

I walked the paper's claimed derivation chain from MEMIT editing (Section 5.2 point 1), through head localization with sigmoid classifiers (Section 5.2 points 2-3), to the MEMAT correction objective (Section 6, Eq. 12). No load-bearing step reduces to its own inputs by construction. The head classifiers are trained to distinguish truthful from false labels on attention-head representations, and the selected head set is then used to optimize additive corrections under a separate likelihood-plus-KL objective; the optimized quantity is not the classifier accuracy itself. The recycled-heads experiments (L1-L2* in Tables 1 and 2, and the scaling curves in Section 6.1) train corrections on one set of factual triplets and apply them to different triplets and languages, which is genuine external validation rather than a fitted prediction. The choice K=16 is made by a hyperparameter search on the evaluation metrics (Appendix F), which is a model-selection concern, not an equation-level circularity. The paper invokes ITI and MEMIT as external prior work but does not rest its central claims on self-citations or imported uniqueness theorems. The lack of a random-head control is a valid experimental-design criticism but does not constitute circularity, since it does not show that any claimed derivation is equivalent to its inputs. Overall, the central empirical claim is self-contained and externally tested, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the base MEMIT editor, the ITI head-selection hypothesis, and the auto-translated CounterFact set. The method introduces no new physical entities; the attention head corrections are trained parameters, not independently motivated postulates.

free parameters (6)
  • K (number of optimized attention heads) = 16
    Selected by hyperparameter search over {8,16,32,48} in Appendix F using the evaluation metrics; no separate held-out criterion reported.
  • lambda_omega (regularization weight in Eq. 12) = not reported
    The L2 penalty on head corrections needs a numeric value; the paper gives the loss form but not the coefficient used.
  • beta (validation fraction used in head scoring, Eq. 11) = not reported
    Controls which validation samples are used to select the top K heads; value not given.
  • MEMIT learning rate = 0.2
    Optimized in Appendix B for Aguila-7B; the base editor MEMAT depends on this.
  • Adam batch size and gradient accumulation steps for head-correction optimization = not reported
    The paper says batch sizes and gradient accumulation are used but does not state their values.
  • Sigmoid classifier regularization C (scikit-learn) = not reported
    Used in the head-locating step; without it the exact head rankings in Section 5.2 cannot be reproduced.
assumptions (6)
  • domain assumption MEMIT is a valid base editor: factual associations are localized in mid-layer MLP weights and can be inserted by solving Eq. 5.
    MEMAT step (a) applies MEMIT unchanged; if MEMIT's edit is not reliable, MEMAT inherits the failure.
  • domain assumption Attention heads linearly separate truthful from false factual associations (ITI hypothesis).
    Section 5.2 extends Li et al. 2023a to post-edit factual associations; the entire head-selection procedure assumes this linear separability.
  • domain assumption Head relevance transfers across languages and to unseen factual associations.
    The method selects heads on a validation split and applies corrections to new facts; if transfer fails, the method's main benefit disappears.
  • domain assumption The automatic English-to-Catalan translation pipeline preserves the factual association structure through subject-relation-object ordering and gender handling.
    Section 4 translates CounterFact with mt-aina-en-ca and simalign, discarding order-shifted cases; translation errors would bias cross-lingual results.
  • domain assumption Aguila-7B is representative enough for cross-lingual knowledge-editing conclusions.
    All experiments use one 6.85B decoder-only model; the Limitations section acknowledges that generalization to other languages and models is untested.
  • domain assumption The purified CounterFact subset is a valid testbed for measuring success, generalization, and specificity.
    Evaluation inherits CounterFact prompts and the Schott et al. purification; results depend on this dataset's construction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mass-Editing Memory with Attention in Transformers: A cross-lingual exploration of knowledge." pith.science (2026). https://pith.science/paper/2I655BOS

@misc{pith2026250202173,
  author       = {Pith},
  title        = {Pith review of: Mass-Editing Memory with Attention in Transformers: A cross-lingual exploration of knowledge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2I655BOS}},
  note         = {Machine review of arXiv:2502.02173}
}
read the original abstract

Recent research has explored methods for updating and modifying factual knowledge in large language models, often focusing on specific multi-layer perceptron blocks. This study expands on this work by examining the effectiveness of existing knowledge editing methods across languages and delving into the role of attention mechanisms in this process. Drawing from the insights gained, we propose Mass-Editing Memory with Attention in Transformers (MEMAT), a method that achieves significant improvements in all metrics while requiring minimal parameter modifications. MEMAT delivers a remarkable 10% increase in magnitude metrics, benefits languages not included in the training data and also demonstrates a high degree of portability. Our code and data are at https://github.com/dtamayo-nlp/MEMAT.

Figures

Figures reproduced from arXiv: 2502.02173 by the authors.

Figure 1
Figure 1. Results of Efficacy, Generalization and Speci [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Accuracy on the validation set for all heads in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration depicting the key steps of MEMAT in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: MEMIT and MEMAT scaling curves plot showing the performance of English and Catalan against number [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Results over training using MEMIT in Catalan [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 5
Figure 5. Figure 5: Example of a Catalan CounterFact sample. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 7
Figure 7. Figure 7: Results over training using PMET in Catalan [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 9
Figure 9. Figure 9: Accuracy on the validation set for all heads in [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 8
Figure 8. Figure 8: Visualization of the frequency distribution of [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 10
Figure 10. Figure 10: Illustration of all the metrics in Catalan (a) or English (b) evaluation when employing MEMAT [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: English Efficacy, Generalization and Specificity general metrics after applying ITI in the context of [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 3 canonical work pages

  1. [1]

    URL: " 'urlintro :=

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

  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]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. http://arxiv.org/abs/1607.06450 Layer normalization

    arXiv 2016

  4. [4]

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. 2022. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, pages 2206--2240. PMLR

    2022

  5. [5]

    Weixin Chen and Bo Li. 2024. Grath: Gradual self-truthifying for large language models. arXiv preprint arXiv:2401.12292

    arXiv 2024

  6. [6]

    Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William W Cohen. 2022 a . Murag: Multimodal retrieval-augmented generator for open question answering over images and text. arXiv preprint arXiv:2210.02928

    arXiv 2022

  7. [7]

    Wenhu Chen, Hexiang Hu, Chitwan Saharia, and William W Cohen. 2022 b . Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491

    arXiv 2022

  8. [8]

    Yuheng Chen, Pengfei Cao, Yubo Chen, Kang Liu, and Jun Zhao. 2023. http://arxiv.org/abs/2308.13198 Journey to the center of the knowledge neurons: Discoveries of language-independent knowledge neurons and degenerate knowledge neurons

    arXiv 2023

Show all 58 references
  1. [9]

    Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. 2023. Dola: Decoding by contrasting layers improves factuality in large language models. arXiv preprint arXiv:2309.03883

    2023 arXiv

  2. [10]

    Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2023. Evaluating the ripple effects of knowledge editing in language models. arXiv preprint arXiv:2307.12976

    2023 arXiv

  3. [11]

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022 a . http://arxiv.org/abs/2104.08696 Knowledge neurons in pretrained transformers

    2022 arXiv

  4. [12]

    Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Zhifang Sui, and Furu Wei. 2022 b . Why can gpt learn in-context? language models secretly perform gradient descent as meta optimizers. arXiv preprint arXiv:2212.10559

    2022 arXiv

  5. [13]

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dari...

    2021

  6. [14]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997

    2023 arXiv

  7. [15]

    Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. 2023. http://arxiv.org/abs/2304.14767 Dissecting recall of factual associations in auto-regressive language models

    2023 arXiv

  8. [16]

    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, ...

    2021 doi

  9. [17]

    Bowman, and Noah A

    Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R. Bowman, and Noah A. Smith. 2018. http://arxiv.org/abs/1803.02324 Annotation artifacts in natural language inference data

    2018 arXiv

  10. [18]

    Danny Halawi, Jean-Stanislas Denain, and Jacob Steinhardt. 2023. http://arxiv.org/abs/2307.09476 Overthinking the truth: Understanding how language models process false demonstrations

    2023 arXiv

  11. [19]

    Thomas Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2022. Aging with grace: Lifelong model editing with discrete key-value adaptors. arXiv preprint arXiv:2211.11031

    2022 arXiv

  12. [20]

    Li, and Jacob Andreas

    Evan Hernandez, Belinda Z. Li, and Jacob Andreas. 2023. http://arxiv.org/abs/2304.00740 Inspecting and editing knowledge representations in language models

    2023 arXiv

  13. [21]

    Robin Jia and Percy Liang. 2017. http://arxiv.org/abs/1707.07328 Adversarial examples for evaluating reading comprehension systems

    2017 arXiv

  14. [22]

    Vladimir Karpukhin, Barlas O g uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906

    2020 arXiv

  15. [23]

    Nora Kassner, Philipp Dufter, and Hinrich Sch \"u tze. 2021. Multilingual lama: Investigating knowledge in multilingual pretrained language models. arXiv preprint arXiv:2102.00894

    2021 arXiv

  16. [24]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. http://arxiv.org/abs/1412.6980 Adam: A method for stochastic optimization

    2017 arXiv

  17. [25]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. http://arxiv.org/abs/2005.11401 Retrieval-augmented generation for knowledge-inte...

    2021 arXiv

  18. [26]

    Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023 a . http://arxiv.org/abs/2306.03341 Inference-time intervention: Eliciting truthful answers from a language model

    2023 arXiv

  19. [27]

    Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2022. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097

    2022 arXiv

  20. [28]

    Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2023 b . http://arxiv.org/abs/2308.08742 Pmet: Precise model editing in a transformer

    2023 arXiv

  21. [29]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. http://arxiv.org/abs/2109.07958 Truthfulqa: Measuring how models mimic human falsehoods

    2022 arXiv

  22. [30]

    Jun-Yu Ma, Jia-Chen Gu, Zhen-Hua Ling, Quan Liu, and Cong Liu. 2023. Untying the reversal curse via bidirectional language model editing. arXiv preprint arXiv:2310.10322

    2023 arXiv

  23. [31]

    Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2020. Generation-augmented retrieval for open-domain question answering. arXiv preprint arXiv:2009.08553

    2020 arXiv

  24. [32]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2023 a . http://arxiv.org/abs/2202.05262 Locating and editing factual associations in gpt

    2023 arXiv

  25. [33]

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2023 b . http://arxiv.org/abs/2210.07229 Mass-editing memory in a transformer

    2023 arXiv

  26. [34]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022 a . http://arxiv.org/abs/2110.11309 Fast model editing at scale

    2022 arXiv

  27. [35]

    Manning, and Chelsea Finn

    Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D. Manning, and Chelsea Finn. 2022 b . http://arxiv.org/abs/2206.06520 Memory-based model editing at scale

    2022 arXiv

  28. [36]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback, 2022. URL https://arxiv. org/abs/2203.02155, 13

    2022 arXiv

  29. [37]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

    2019 arXiv

  30. [38]

    Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Andreas Müller, Joel Nothman, Gilles Louppe, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu B...

    2018 arXiv

  31. [39]

    Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. 2023. http://arxiv.org/abs/2306.01116 The refinedweb dataset for falcon llm: Outperforming curated corpora with...

    2023 arXiv

  32. [40]

    Yuval Pinter and Michael Elhadad. 2023. Emptying the ocean with a spoon: Should we edit models? arXiv preprint arXiv:2310.11958

    2023 arXiv

  33. [41]

    Jirui Qi, Raquel Fern \'a ndez, and Arianna Bisazza. 2023. Cross-lingual consistency of factual knowledge in multilingual language models. arXiv preprint arXiv:2310.10378

    2023 arXiv

  34. [42]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290

    2023 arXiv

  35. [43]

    Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

    2009

  36. [44]

    Masoud Jalili Sabet, Philipp Dufter, François Yvon, and Hinrich Schütze. 2021. http://arxiv.org/abs/2004.08728 Simalign: High quality word alignments without parallel training data using static and contextualized embeddings

    2021 arXiv

  37. [45]

    Mansi Sakarvadia, Aswathy Ajith, Arham Khan, Daniel Grzenda, Nathaniel Hudson, Andr \'e Bauer, Kyle Chard, and Ian Foster. 2023. Memory injections: Correcting multi-hop reasoning failures during inference in transformer-based language models. arXiv preprint arXiv:2309.05605

    2023 arXiv

  38. [46]

    Tim Schott, Daniel Furman, and Shreshta Bhat. 2023. http://arxiv.org/abs/2305.13675 Polyglot or not? measuring multilingual encyclopedic knowledge retrieval from foundation language models

    2023 arXiv

  39. [47]

    Noam Shazeer. 2019. http://arxiv.org/abs/1911.02150 Fast transformer decoding: One write-head is all you need

    2019 arXiv

  40. [48]

    Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. 2020. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008--3021

    2020

  41. [49]

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. 2023. http://arxiv.org/abs/2104.09864 Roformer: Enhanced transformer with rotary position embedding

    2023 arXiv

  42. [50]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. http://arxiv.org/abs/1706.03762 Attention is all you need

    2017 arXiv

  43. [51]

    Changhan Wang, Kyunghyun Cho, and Jiatao Gu. 2020. Neural machine translation with byte-level subwords. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9154--9160

    2020

  44. [52]

    Jiaan Wang, Yunlong Liang, Zengkui Sun, Yuxuan Cao, and Jiarong Xu. 2023 a . http://arxiv.org/abs/2309.08952 Cross-lingual knowledge editing in large language models

    2023 arXiv

  45. [53]

    Weixuan Wang, Barry Haddow, and Alexandra Birch. 2023 b . Retrieval-augmented multilingual knowledge editing. arXiv preprint arXiv:2312.13040

    2023 arXiv

  46. [54]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020 arXiv

  47. [55]

    Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2022. Retrieval-augmented multimodal language modeling. arXiv preprint arXiv:2211.12561

    2022 arXiv

  48. [56]

    Mert Yuksekgonul, Varun Chandrasekaran, Erik Jones, Suriya Gunasekar, Ranjita Naik, Hamid Palangi, Ece Kamar, and Besmira Nushi. 2023. Attention satisfies: A constraint-satisfaction lens on factual errors of language models. arXiv preprint arXiv:2309.15098

    2023 arXiv

  49. [57]

    Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023. http://arxiv.org/abs/2305.12740 Can we edit factual knowledge by in-context learning?

    2023 arXiv

  50. [58]

    Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363

    2020 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.