MEMIT-Merge: Addressing MEMIT's Key-Value Conflicts in Same-Subject Batch Editing for LLMs

Rui Xia; Xiangqing Shen; Zilu Dong

REVIEW 3 major objections 7 minor 1 cited by

MEMIT-Merge: Addressing MEMIT's Key-Value Conflicts in Same-Subject Batch Editing for LLMs

T0 review · 3 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read MEMIT-Merge claims that a single merged value per subject resolves MEMIT's key-value collisions, raising same-subject edit success from roughly 41% to 95%.

desk verdict A real and previously undocumented failure mode of MEMIT, plus a simple fix that works for small batches; the abstract's >90% claim is overstated once batch size exceeds roughly 30. read the letter →

arxiv 2502.07322 v3 pith:RPFTB2J3 submitted 2025-02-11 cs.CL cs.LG

Zilu Dong , Xiangqing Shen , Rui Xia This is my paper · ORCID

classification cs.CLcs.LG

keywords knowledgeeditingMEMITbatchsame-subjecteditskey-valueconflictMLPmemorycounterfactualdatasetslargelanguagemodels

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

MEMIT, a widely used method for writing many facts into a language model in one update, loses most of its accuracy when a batch contains several edits about the same person or entity. The paper traces this to a key-value conflict inside the MLP (the transformer's feed-forward memory layers): the key computed at the subject's last token is essentially the same for all the edits, yet each edit demands a different output value. MEMIT-Merge removes the conflict by optimizing one merged value for all facts that share a subject, instead of one value per fact. On ten same-subject edits, the reported success rate rises from about 0.41 to 0.95 on Qwen2.5-1.5B and from 0.51 to 1.00 on Llama-3-8B, while distinct-subject performance is unchanged. If this holds, batch editing becomes practical for realistic updates such as refreshing several attributes of one person at once.

What carries the argument

The load-bearing object is the merged value vector $v$ computed by Eq. (3): for the set of edited facts sharing the same subject, it maximizes the summed log-probability of all the target objects, then feeds that single value into the same key-value update MEMIT uses for the MLP output weight. This converts a batch of contradictory key-value pairs into one consistent pair per subject. The price is that one vector must store every relation-object association for that subject.

What would settle it

Take the same-subject batch of size 10, compute the merged value $v$ from Eq. (3), and apply the MEMIT update; then check each edited sentence's completion probability. If any requested object is not the most likely completion under the edited model—or if the summed probability of all target objects is no higher than before editing—the claim that one merged value can serve all facts for a subject is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that MEMIT's batch edit capacity is governed by key identity rather than batch size: when edits share a subject, the keys at the subject's last token collide while the values optimized for each object diverge, forcing the same matrix to map one input to many outputs. MEMIT-Merge replaces the per-fact value objective with a joint objective over all relation-object pairs that share the key, so each subject contributes exactly one consistent key-value pair to the parameter update. The paper supports this mechanism with a same-subject versus distinct-subject dataset pair and an average-key-distance analysis showing that nearer keys produce more severe degradation.

Load-bearing premise

The method assumes a single merged value vector can simultaneously represent all the relation-object pairs that a subject must point to, even though the key at the subject's last token contains no relation information and the paper acknowledges the value vector's capacity is limited.

Editorial extensions

If this is right

Same-subject batch editing becomes usable for realistic loads: updating a person's occupation, workplace, and family members in one edit pass no longer collapses to the unedited model.
Because efficacy degrades with smaller key distance, MEMIT-style editors should be expected to fail on semantically or lexically similar subjects, not only on identical strings; the average-key-distance metric gives a way to predict when.
The merge is a value-computation change, so it can be attached to MEMIT-derived methods; the paper shows a PMET-Merge variant also mitigates the drop, though less strongly.
Distinct-subject performance and specificity are preserved, so applying MEMIT-Merge does not trade away the original method's strengths on standard batches.

Reading between the lines

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

Editorial inference: any locate-and-edit method that derives its key from the subject's last token should show the same collision, so key-wise merging is a generic patch for that family, not specific to MEMIT.
Editorial inference: because the paper itself identifies the relation-free key as the root cause, the natural next step is to edit at a token that encodes the relation, for instance the object token; MEMIT-Merge is a workaround, not the final architecture.
Editorial inference: the reported same-subject dataset uses only person subjects and lexical identity, so an immediate test is whether merging transfers to places and organizations, or to subjects that are distinct strings but close in embedding space; the AKD metric supplies a quantitative handle.
Editorial inference: the value-capacity ceiling implies an information-theoretic limit: one vector can represent only so many associations, so very large same-subject batches will need multiple keys per subject or a sparser update scheme.

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.

Desk Editor's Note

read the letter

Colleague,

You should know two things about this paper. First, it identifies a genuine, previously undocumented failure mode of MEMIT: when a batch contains multiple edits sharing the same subject, the edit success rate collapses because identical keys are forced to map to conflicting values. Second, the proposed fix—grouping same-subject facts and optimizing a single merged value (Eq. 3)—works well for small batches, but the abstract's claim of "maintaining >90%" only holds up to about batch 30. The authors admit this in the Limitations section: a single value vector cannot represent unlimited facts, and the key at the subject's last token carries no relation information.

What's genuinely new is the diagnosis of the key-collision problem and the AKD correlation analysis. The evidence is consistent across four models (Qwen2.5-1.5B, Qwen2.5-7B, Llama-3-8B, GPT-J-6B). At batch size 10, MEMIT-Merge lifts same-subject efficacy from around 0.34–0.51 to 0.95–1.00 while leaving distinct-subject performance unchanged. The fix is simple enough to drop into existing MEMIT pipelines, and the code is available.

Soft spots, in proportion. The main weakness is scope. The experiments use one constructed dataset with the single fixed subject string "John Smith", 100 hand-filtered person relations, and no error bars or repeated shuffles. The value-capacity ceiling is real and quantified in Fig. 3: efficacy falls below 90% once the batch exceeds roughly 30, and the authors acknowledge this. So the paper's headline claim is a small-batch statement, not a general robustness result. Also, the dataset is not released (only code), and generalization to non-person subjects and semantic similarity in embedding space is untested, as the Limitations honestly state.

The central argument holds up: the key-collision mechanism is mathematically plausible and supported by the AKD correlation and by the FT baseline showing that non-MEMIT methods don't suffer the same degradation. This is a scope restriction, not a refutation.

Who should read it: anyone working on mass knowledge editing, especially practical applications like updating a person's profile. It deserves a serious referee. The paper needs revision to temper the abstract's claim, add error bars and more diverse subject types, and release the dataset. But the diagnosis is valuable and the fix is practical.

My recommendation: send it to peer review, expecting a revise-and-resubmit.

Referee Report

3 major / 7 minor

Summary. The paper identifies a failure mode of MEMIT batch knowledge editing: when multiple edits in a batch share the same subject, MEMIT's efficacy drops sharply because the subject-derived key is identical while the optimized values differ. The authors propose MEMIT-Merge, which optimizes a single merged value for all facts sharing a key (Eq. 3) and then applies the standard MEMIT closed-form update. Experiments on four models with custom same-subject and distinct-subject datasets show that at batch size 10 the merged variant restores efficacy to roughly 0.95-1.00 where MEMIT is between 0.34 and 0.51, while distinct-subject performance is essentially unchanged. The paper also presents an AKD analysis linking key distance to editing efficacy and includes a limitations section acknowledging the finite capacity of a single value vector.

Significance. If the result holds, the paper identifies a real and practically important limitation of a widely used editing algorithm and offers a simple, parameter-free fix. The strengths are that the same-subject degradation is replicated across multiple model architectures, the fix consistently improves same-subject efficacy at batch size 10, the FT baseline provides a meaningful control, and the code is released. The AKD analysis is a useful independent diagnostic. However, the headline '>90% robustness' is only established for small batches and for a single fixed subject string, and no error bars or repeated runs are reported. With appropriate scope qualifications and variance estimates, the contribution would be solid and of clear interest to the knowledge-editing community.

major comments (3)

[Abstract / Sec. 4.4 / Fig. 3] The abstract and introduction claim that when MEMIT's efficacy drops to about 50%, MEMIT-Merge 'maintains a success rate exceeding 90%,' but this holds only for small batches. In Fig. 3, the same-subject efficacy of MEMIT-Merge on Qwen2.5-1.5B falls below 0.9 once the batch size exceeds roughly 30 and is around 0.6 at batch size 100. Because the paper's own Sec. 4.4 and Limitations attribute this to the limited capacity of a single merged value vector, the central robustness claim should be qualified by batch size (e.g., 'for batches of up to about 30 same-subject edits') in the abstract and conclusion.
[Tabs. 1 and 3, Figs. 3 and 6] All same-subject results are produced from a single fixed subject string ('John Smith') and one hand-filtered set of 100 relations, with no repeated runs, no shuffles, and no error bars. Since both the severity of the original failure and the capacity ceiling of the merged value can depend on the specific subject and on the ordering of relations, the headline numbers (e.g., 0.41 vs 0.95 for Qwen2.5-1.5B, 0.51 vs 1.00 for Llama-3-8B) may reflect a favorable draw. Please report results over multiple subjects and multiple shuffles, or at least provide standard errors or confidence intervals for the batch-size-10 numbers.
[Sec. 3.2] The statement that MEMIT-Merge has 'superior theoretical efficiency' because merging 'effectively increases the batch size during gradient descent' is not supported: a larger gradient batch requires more forward and backward computation per optimized value, not less. If the intended efficiency gain is the reduction in the number of (key, value) pairs fed to the closed-form update in Eq. (2), that effect should be stated explicitly and quantified, or the sentence should be removed.

minor comments (7)

[Abstract] The abstract contains a duplicated sentence: 'Addressing this issue, we propose MEMIT-Merge... effectively resolving the performance degradation in same-subject batch editing scenarios.' appears twice in a row.
[Table 3] The table headers read 'same-dataset' and 'distinct-dataset' rather than 'same-subject' and 'distinct-subject'; please fix for consistency with the rest of the paper.
[Sec. 4.2] Section 4.2 states that experiments use three models, but Appendices D and E report results also for Qwen2.5-7B-Instruct, making four models total; please correct the count.
[Sec. 4.2 and tables] The fine-tuning baseline is called FT-L in Sec. 4.2 but is labeled FT in all tables; use one name consistently.
[Sec. 4.3] The 'anomalously high specificity' of MEMIT on same-subject edits is not anomalous if the edit failed: high specificity simply means the model was unchanged. Please rephrase to avoid implying that high specificity is itself a defect.
[Appendix E] The text says the same phenomenon observed in GPT-J also exists in Qwen2.5 and Llama-3, but Fig. 6 does not include GPT-J; the reference to GPT-J should be clarified.
[Appendix B / Fig. 4] The AKD correlation is presented as 'statistically significant' without any test; please either add a statistical test or soften the wording to 'consistent with a negative correlation'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MEMIT-Merge's Eq. 3 is a new joint optimization objective rather than a restatement of MEMIT's inputs, and the reported efficacy gains are empirical benchmark results with acknowledged scope limits.

full rationale

The paper's central diagnostic claim—that same-subject batches force identical keys to map to conflicting values—is a direct consequence of MEMIT's key-value formalism and the paper's dataset construction, not a conclusion derived from the paper's own outputs. The value objective in Eq. 3 replaces MEMIT's per-fact argmin with a sum over facts sharing a key; it is a proposed algorithm, not a fitted parameter renamed as a prediction. The >90% same-subject efficacy is an empirical outcome of running this optimizer on Qwen2.5, Llama-3, and GPT-J, and the paper itself shows the result degrades at larger batch sizes (Fig. 3), so the result is not tautologically forced by the objective. The key-collision mechanism is also supported by an independent correlational AKD analysis rather than by assuming the conclusion. There are no load-bearing self-citations: the paper builds on MEMIT, PMET, and AlphaEdit as external prior work by other authors, and no uniqueness theorem or ansatz is imported from the present authors' earlier work. The limitations (single value vector capacity is not unlimited; the key at the subject's last token carries no relation information) are explicitly acknowledged scope constraints, not circular moves. While the same datasets motivate both the diagnosis and the fix, that is standard experimental design rather than circularity, and no equation in the paper reduces to its own input by construction.

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

No new fitted parameters are introduced; MEMIT-Merge changes only how the value vector is optimized. The load-bearing assumptions are inherited from the MEMIT framework and the paper's own claim that one value vector can represent several facts, which the authors acknowledge is limited.

assumptions (4)

domain assumption Factual knowledge in LLMs is stored in MLP key-value associations, and editing W_out in the MLP can insert facts.
Inherited from ROME/MEMIT (Sec. 2.1, Sec. 3.1). If this storage hypothesis is wrong, the whole locate-and-edit framework, including the proposed fix, is moot.
domain assumption All edits sharing the same subject produce identical or near-identical keys at the editable layer.
Used in Sec. 3.1 and Fig. 2 to define the key collision problem; the AKD analysis in App. B is the only empirical support, showing AKD=0 for exact same-subject batches.
domain assumption A single optimized value vector v can satisfy multiple same-subject facts simultaneously (Eq. 3 has a usable solution).
This is the core design premise of MEMIT-Merge. The paper acknowledges it is limited because the key lacks relation information and a value vector has finite capacity (Limitations).
standard math The closed-form update Eq. (2) correctly applies merged values to the model without destroying unrelated knowledge.
Ridge-regression solution inherited from MEMIT; standard linear algebra assuming C + KK^T is invertible.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MEMIT-Merge: Addressing MEMIT's Key-Value Conflicts in Same-Subject Batch Editing for LLMs." pith.science (2026). https://pith.science/paper/RPFTB2J3

@misc{pith2026250207322,
  author       = {Pith},
  title        = {Pith review of: MEMIT-Merge: Addressing MEMIT's Key-Value Conflicts in Same-Subject Batch Editing for LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RPFTB2J3}},
  note         = {Machine review of arXiv:2502.07322}
}

read the original abstract

As large language models continue to scale up, knowledge editing techniques that modify models' internal knowledge without full retraining have gained significant attention. MEMIT, a prominent batch editing algorithm, stands out for its capability to perform mass knowledge modifications. However, we uncover that MEMIT's editing efficacy significantly deteriorates when processing batches containing multiple edits sharing the same subject. Our analysis reveals this stems from MEMIT's key value modeling framework: identical keys (derived from the shared subject) are forced to represent different values (corresponding to different knowledge), resulting in update conflicts during editing. Addressing this issue, we propose MEMIT-Merge, an enhanced approach that merges value computation processes for facts sharing the same subject, effectively resolving the performance degradation in samesubject batch editing scenarios. Experimental results demonstrate that when MEMIT's edit success rate drops to around 50% at larger batch sizes, MEMIT-Merge maintains a success rate exceeding 90%, showcasing remarkable robustness to subject entity collisions. The code is available at https://github.com/NUSTM/ MEMIT-Merge.

Figures

Figures reproduced from arXiv: 2502.07322 by the authors.

**Figure 2.** The architecture of MEMIT processing two same subject sentences. The left and right sides of the figure [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The results of MEMIT-Merge and MEMIT methods on same-subject and distinct-subject datasets using the Qwen2.5-1.5B-Instruct model, showing the changes with varying batch sizes. MEMIT-Merge is capable of significantly alleviating the decline in editing performance under the same-subject condition. the standard edit success rate plummets at a batch size of 2, whereas MEMIT-Merge is able to maintain a much higher succe… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Datasets with different AKD values and the results of edit efficacy. The lower the AKD value, the more severe the decline in edit capability. have a relatively consistent distance between each other. We utilize the knowledge triples from the samesubject and distinct-s…

**Figure 5.** Figure 5: Editing same-subject dataset using Qwen2.5- [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

ChainEdit: Propagating Ripple Effects in LLM Knowledge Editing through Logical Rule-Guided Chains
cs.CL 2025-07 conditional novelty 5.0 of 10

ChainEdit uses knowledge-graph logical rules, filtered by an LLM, to propagate a single fact edit to dependent facts and lifts logical generalization on RIPPLE EDITS by about 40 points.

Reference graph

Works this paper leans on

17 extracted references · 3 canonical work pages · cited by 1 Pith paper

[1]

AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

2024
[2]

Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022. https://doi.org/10.18653/V1/2022.ACL-LONG.581 Knowledge neurons in pretrained transformers . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022 , pages 8493--8502. Associat...

doi:10.18653/v1/2022.acl-long.581 2022
[3]

Jingcheng Deng, Zihao Wei, Liang Pang, Hanxing Ding, Huawei Shen, and Xueqi Cheng. 2024. https://doi.org/10.48550/arXiv.2405.15349 UnKE : Unstructured Knowledge Editing in Large Language Models . arXiv preprint. ArXiv:2405.15349 [cs]

arXiv doi:10.48550/arxiv.2405.15349 2024
[4]

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Jie Shi, Xiang Wang, Xiangnan He, and Tat - Seng Chua. 2025. https://openreview.net/forum?id=HvSytvg3Jh Alphaedit: Null-space constrained knowledge editing for language models . In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net

work page 2025
[5]

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Xiang Wang, Xiangnan He, and Tat-seng Chua. 2024. https://doi.org/10.48550/arXiv.2410.02355 AlphaEdit : Null - Space Constrained Knowledge Editing for Language Models . arXiv preprint. ArXiv:2410.02355

arXiv doi:10.48550/arxiv.2410.02355 2024
[6]

Akshat Gupta, Dev Sajnani, and Gopala Anumanchipalli. 2024. https://aclanthology.org/2024.findings-emnlp.903 A unified framework for model editing . In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024 , pages 15403--15418. Association for Computational Linguistics

work page 2024
[7]

Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2024. https://doi.org/10.1609/AAAI.V38I17.29818 PMET: precise model editing in a transformer . In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educat...

doi:10.1609/aaai.v38i17.29818 2024
[8]

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. http://papers.nips.cc/paper\_files/paper/2022/hash/6f1d43d5a82a37e89b0665b33bf3a182-Abstract-Conference.html Locating and editing factual associations in GPT . In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2...

2022

Show all 17 references

[9]

Andonian, Yonatan Belinkov, and David Bau

Kevin Meng, Arnab Sen Sharma, Alex J. Andonian, Yonatan Belinkov, and David Bau. 2023. https://openreview.net/forum?id=MkbcAHIYgyS Mass-editing memory in a transformer . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2...

2023
[10]

Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022. https://openreview.net/forum?id=0DcZxeWfOPt Fast model editing at scale . In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 202...

2022
[11]

Qwen, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le ...

2025 doi arXiv
[12]

Chenmien Tan, Ge Zhang, and Jie Fu. 2024. https://openreview.net/forum?id=L6L1CJQ2PE Massive editing for large language models via meta learning . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net

2024
[13]

Ben Wang and Aran Komatsuzaki. 2021. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model . https://github.com/kingoflolz/mesh-transformer-jax

2021
[14]

Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, Siyuan Cheng, Ziwen Xu, Xin Xu, Jia-Chen Gu, Yong Jiang, Pengjun Xie, Fei Huang, Lei Liang, Zhiqiang Zhang, Xiaowei Zhu, Jun Zhou, and Huajun Chen. ...

2024 doi arXiv
[15]

Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. https://doi.org/10.48550/arXiv.2012.00363 Modifying Memories in Transformer Models . arXiv preprint. ArXiv:2012.00363

2020 doi arXiv
[16]

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...
[17]

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 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

2024

[2] [2]

Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022. https://doi.org/10.18653/V1/2022.ACL-LONG.581 Knowledge neurons in pretrained transformers . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022 , pages 8493--8502. Associat...

doi:10.18653/v1/2022.acl-long.581 2022

[3] [3]

Jingcheng Deng, Zihao Wei, Liang Pang, Hanxing Ding, Huawei Shen, and Xueqi Cheng. 2024. https://doi.org/10.48550/arXiv.2405.15349 UnKE : Unstructured Knowledge Editing in Large Language Models . arXiv preprint. ArXiv:2405.15349 [cs]

arXiv doi:10.48550/arxiv.2405.15349 2024

[4] [4]

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Jie Shi, Xiang Wang, Xiangnan He, and Tat - Seng Chua. 2025. https://openreview.net/forum?id=HvSytvg3Jh Alphaedit: Null-space constrained knowledge editing for language models . In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net

work page 2025

[5] [5]

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Xiang Wang, Xiangnan He, and Tat-seng Chua. 2024. https://doi.org/10.48550/arXiv.2410.02355 AlphaEdit : Null - Space Constrained Knowledge Editing for Language Models . arXiv preprint. ArXiv:2410.02355

arXiv doi:10.48550/arxiv.2410.02355 2024

[6] [6]

Akshat Gupta, Dev Sajnani, and Gopala Anumanchipalli. 2024. https://aclanthology.org/2024.findings-emnlp.903 A unified framework for model editing . In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024 , pages 15403--15418. Association for Computational Linguistics

work page 2024

[7] [7]

Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2024. https://doi.org/10.1609/AAAI.V38I17.29818 PMET: precise model editing in a transformer . In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educat...

doi:10.1609/aaai.v38i17.29818 2024

[8] [8]

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. http://papers.nips.cc/paper\_files/paper/2022/hash/6f1d43d5a82a37e89b0665b33bf3a182-Abstract-Conference.html Locating and editing factual associations in GPT . In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2...

2022