When to Write and When to Suppress: Route-Specialized Dual Adapters for Memory-Assisted Knowledge Editing

Baijia Zhang; Yining Huang

arxiv: 2606.14668 · v3 · pith:H5AUZKONnew · submitted 2026-06-12 · 💻 cs.LG

When to Write and When to Suppress: Route-Specialized Dual Adapters for Memory-Assisted Knowledge Editing

Baijia Zhang , Yining Huang This is my paper

Pith reviewed 2026-06-27 04:43 UTC · model grok-4.3

classification 💻 cs.LG

keywords knowledge editingmemory-assisted editingrelevance routingdual adaptersparameter-efficient fine-tuninglocality preservationLLM editing

0 comments

The pith

A relevance router pairs with dual adapters to decide when to apply or suppress memory edits in language models.

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

The paper proposes that knowledge editing benefits from explicitly choosing when to write a new fact versus when to suppress an edit. It trains one adapter to shift object preference toward the edit and a second adapter to preserve the original preference on non-routed prompts. A router first classifies whether a prompt should access the edit memory. On Llama-3.1-8B-Instruct this yields the highest probability-preference accuracy across three 1,000-case benchmarks. Ablations indicate the improvement stems from separating the two behaviors rather than from added adapter capacity alone.

Core claim

Route-specialized dual adapters let a model retrieve an edit memory at inference time, then apply either an edit adapter that prefers the new object or a locality adapter that restores the original-object preference, depending on a relevance router decision.

What carries the argument

Relevance router that routes prompts to either an edit adapter or a locality adapter, each trained separately under the same memory protocol.

If this is right

Lexical neural routing performs best on CF while embedding routing performs best on ZSRE and MQuAKE.
The dual-adapter split produces higher overall accuracy than increasing LoRA rank or capacity in a single adapter.
The same router-plus-dual-adapter pattern improves results on both Llama-3.1-8B-Instruct and Qwen3-8B.
Edit success and locality can be optimized independently by training the two adapters on disjoint prompt sets.

Where Pith is reading between the lines

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

The approach could be tested on sequential editing streams where new facts arrive over time and router decisions must remain stable.
Router training might be made dataset-agnostic by combining lexical and embedding signals in one model.
The separation of edit and suppression objectives may reduce unintended changes when multiple overlapping edits are stored in the same memory.

Load-bearing premise

The relevance router correctly separates prompts that need an edit from those that should keep the original behavior, and the two adapters do not interfere when both are loaded.

What would settle it

Measuring whether accuracy falls below single-adapter baselines on any benchmark when the router is replaced by random routing or when both adapters are merged into one.

Figures

Figures reproduced from arXiv: 2606.14668 by Baijia Zhang, Yining Huang.

**Figure 1.** Figure 1: Route-specialized memory-assisted editing. A relevance router decides whether to retrieve edit memory. Routed prompts use the edit adapter and are optimized to prefer the new object. Unrouted non-direct prompts use the locality adapter and are optimized to prefer the protected original object. where Lrank = softplus(m − score(x, y+) + score(x, y−)). (4) For direct and paraphrase examples, (y +, y−) = (o ⋆ … view at source ↗

**Figure 2.** Figure 2: Overall probability-preference accuracy on two base models. RRDA is best on all three benchmarks for both Llama-3.1-8B and Qwen3-8B. CF, contributes nontrivially to direct and paraphrase preferences. Thus the main gain comes from explicit memory, while the dual-adapter route policy improves the reliability-locality balance needed for the full result. 4.4 Router Ablation [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗

**Figure 3.** Figure 3: Different relevance boundaries are needed across benchmarks. CounterFact selects NN because it preserves no-route locality, while ZsRE and MQuAKE-CF select BGE semantic routing. applications where edit provenance matters and where silent probability leakage is unacceptable. 6 Limitations There are several important limitations. First, the baselines are resource-matched PEFT-style implementations, not offi… view at source ↗

**Figure 4.** Figure 4: Locality in the component ablation. The locality adapter is most informative on CounterFact and ZsRE, where off-route prompts are present and nontrivial [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 5.** Figure 5: More target modules do not materially improve CounterFact accuracy under the rank-8, 500-step setting. 2. De Cao, N., Aziz, W., Titov, I.: Editing factual knowledge in language models. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 6491–6506 (2021). https://doi.org/10.18653/v1/2021.emnlpmain.522 3. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: … view at source ↗

read the original abstract

Knowledge editing systems must update selected facts while preserving nearby but irrelevant behavior. This paper studies this problem in a memory-assisted setting where an edit memory is retrieved at inference time and a parameter-efficient adapter corrects the model's object preference. We argue that the central design question is not only how to write an edit, but also when to suppress it. We introduce \method{}, a route-specialized dual-adapter editor. A relevance router first decides whether a prompt should receive an edit memory. Routed prompts use an edit adapter trained to prefer the new object over the original object; unrouted non-direct prompts use a separate locality adapter trained to preserve or restore the original-object preference. We evaluate \method{} on three 1,000-case protocols, \cf{}, \zsre{}, and \mquake{}, under the same memory protocol and two 7B/8B base models. On Llama-3.1-8B-Instruct, \method{} obtains the best overall probability-preference accuracy on all three benchmarks: 0.8180 on \cf{}, 0.8946 on \zsre{}, and 0.9922 on \mquake{}. The same trend holds on Qwen3-8B. Router ablations show that the relevant memory boundary differs across datasets: a lexical neural router is safest on \cf{}, while BGE embedding routing is better on \zsre{} and \mquake{}. Component and module ablations show that the gain mainly comes from separating edit injection from off-route suppression rather than from simply increasing LoRA capacity.

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.

Desk Editor's Note private letter to a colleague

Dual adapters plus a router separate edit injection from locality suppression and deliver the top scores on three standard benchmarks.

read the letter

The core idea is straightforward: a relevance router picks whether a prompt should hit an edit adapter (trained to prefer the new fact) or a separate locality adapter (trained to keep the original preference). This split is presented as the main source of gains rather than just adding more adapter capacity.

The paper shows the approach beats prior single-adapter and memory baselines on CF, ZSRE, and MQuAKE for both Llama-3.1-8B-Instruct and Qwen3-8B, with the largest margins on the harder protocols. Router ablations are included and confirm that the best router type changes with the dataset, which they treat as an empirical finding rather than a hidden variable.

The main limitation is that router selection remains dataset-dependent, so deployment would require some validation step per new domain. The memory retrieval protocol itself follows existing work, so the novelty sits mainly in the dual-adapter routing rather than in the memory mechanism.

This is useful reading for anyone already running knowledge-editing experiments on 7-8B models. The experiments are concrete, the ablations address the central design claim, and the numbers are reported across two base models, so the paper is worth sending to referees.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Route-Specialized Dual Adapters (<method>), a memory-assisted knowledge editing approach in which a relevance router first classifies whether a prompt should receive an edit memory. Routed prompts are handled by an edit adapter trained to prefer the new object; non-routed prompts use a separate locality adapter trained to preserve the original-object preference. The method is evaluated under a fixed memory protocol on the CF, ZSRE, and MQuAKE benchmarks (1,000 cases each) using Llama-3.1-8B-Instruct and Qwen3-8B, reporting the highest probability-preference accuracies of 0.8180, 0.8946, and 0.9922 on the Llama model. Router ablations show dataset-specific optimal routers, and component ablations attribute gains to the separation of edit injection from suppression rather than added LoRA capacity.

Significance. If the reported accuracies and ablation attributions hold under detailed verification, the work offers a clear advance in knowledge editing by explicitly modeling the decision of when to suppress an edit. The empirical superiority across three standard benchmarks, the demonstration that dual specialization outperforms capacity increases, and the dataset-specific router analysis provide both a practical method and useful design insight for parameter-efficient editing systems.

major comments (2)

[Router ablations] Router ablations: The relevance router is load-bearing for the central claim, yet the manuscript reports only that optimal router type differs by dataset (lexical neural safest on CF; BGE embedding better on ZSRE and MQuAKE) without providing router accuracy metrics, misrouting error analysis, or testing a single router across all datasets. This leaves the weakest assumption—that the router accurately identifies edit versus suppression cases—unverified.
[Component ablations] Component ablations: The claim that performance gains derive from separating edit injection from off-route suppression rather than increased LoRA capacity is central, but the manuscript does not report exact parameter counts, training configurations, or baseline single-adapter setups used in the comparison, making the attribution difficult to assess.

minor comments (2)

[Abstract] The abstract states that results hold 'under the same memory protocol' but provides no explicit description of that protocol, which would aid reproducibility and verification of the memory-assisted setting.
[Evaluation] Reported accuracies lack any indication of variance, number of runs, or statistical significance testing, which would strengthen the 'best overall' claim across the three benchmarks.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The two major comments identify gaps in the router and component ablation sections that we will address directly in revision.

read point-by-point responses

Referee: [Router ablations] Router ablations: The relevance router is load-bearing for the central claim, yet the manuscript reports only that optimal router type differs by dataset (lexical neural safest on CF; BGE embedding better on ZSRE and MQuAKE) without providing router accuracy metrics, misrouting error analysis, or testing a single router across all datasets. This leaves the weakest assumption—that the router accurately identifies edit versus suppression cases—unverified.

Authors: We agree that router accuracy metrics and misrouting analysis are needed to substantiate the central claim. In the revised manuscript we will report per-dataset router accuracy (edit vs. non-edit classification) on held-out prompts, include a confusion-matrix style error analysis of misrouted cases, and add results for a single fixed router (the best-performing one) evaluated across all three benchmarks to demonstrate cross-dataset robustness. revision: yes
Referee: [Component ablations] Component ablations: The claim that performance gains derive from separating edit injection from off-route suppression rather than increased LoRA capacity is central, but the manuscript does not report exact parameter counts, training configurations, or baseline single-adapter setups used in the comparison, making the attribution difficult to assess.

Authors: We will expand the component ablation section to include exact trainable parameter counts for the dual-adapter and single-adapter configurations, full training hyper-parameters (learning rate, epochs, batch size), and a clearer description of the single-adapter baselines (identical total capacity, same training data mixture). These additions will allow readers to verify that gains are attributable to specialization rather than capacity. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method validated on external benchmarks

full rationale

The paper proposes a dual-adapter architecture with a relevance router for knowledge editing and evaluates it through direct accuracy measurements and ablations on three standard benchmarks (CF, ZSRE, MQuAKE) using Llama-3.1-8B-Instruct and Qwen3-8B. All reported gains are attributed to component separation via explicit router and module ablations rather than any closed-form derivation, fitted parameter renamed as prediction, or self-citation chain. No equations, uniqueness theorems, or ansatzes are invoked that reduce to the method's own inputs by construction. The work is self-contained against external benchmarks and does not rely on internal self-referential logic for its central claims.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

Paper relies on standard assumptions of adapter training and benchmark evaluation protocols; no new free parameters, axioms, or invented entities are identifiable from the abstract alone.

free parameters (1)

router selection per dataset
Choice between lexical neural router and BGE embedding routing is dataset-dependent and appears tuned rather than derived.

pith-pipeline@v0.9.1-grok · 5824 in / 975 out tokens · 30004 ms · 2026-06-27T04:43:40.990122+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 7 canonical work pages

[1]

Transac- tions of the Association for Computational Linguistics12, 283–298 (2024)

Cohen, R., Biran, E., Yoran, O., Globerson, A., Geva, M.: Evaluat- ing the ripple effects of knowledge editing in language models. Transac- tions of the Association for Computational Linguistics12, 283–298 (2024). https://doi.org/10.1162/tacl_a_00644 Route-Specialized Dual Adapters 13 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Trainable parameters (millions) 0....

work page doi:10.1162/tacl_a_00644 2024
[2]

In: Proceedings of the 2021 Confer- ence on Empirical Methods in Natural Language Processing

De Cao, N., Aziz, W., Titov, I.: Editing factual knowledge in language models. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Lan- guage Processing. pp. 6491–6506 (2021). https://doi.org/10.18653/v1/2021.emnlp- main.522

work page doi:10.18653/v1/2021.emnlp- 2021
[3]

In: Advances in Neural Information Processing Systems

Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient fine- tuning of quantized LLMs. In: Advances in Neural Information Processing Systems. vol. 36 (2023)

2023
[4]

arXiv preprint arXiv:2407.21783 (2024)

Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The Llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

Pith/arXiv arXiv 2024
[5]

In: In- ternational Conference on Learning Representations (2025)

Fang, J., Jiang, H., Wang, K., Ma, Y., Shi, J., Wang, X., He, X., Chua, T.S.: AlphaEdit: Null-space constrained knowledge editing for language models. In: In- ternational Conference on Learning Representations (2025)

2025
[6]

arXiv preprint arXiv:2406.04496 (2024)

Ge, X., Mousavi, A., Grave, E., Joulin, A., Qian, K., Han, B., Arefiyan, M., Li, Y.: Time sensitive knowledge editing through efficient finetuning. arXiv preprint arXiv:2406.04496 (2024)

arXiv 2024
[7]

arXiv preprint arXiv:2502.19416 (2025)

Gupta, A., Fang, C., Ozdemir, A., Lu, M., Alaa, A., Hartvigsen, T., Anu- manchipalli, G.: Norm growth and stability challenges in localized sequential knowledge editing. arXiv preprint arXiv:2502.19416 (2025)

arXiv 2025
[8]

arXiv preprint arXiv:2211.11031 (2023)

Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., Ghassemi, M.: Ag- ing with GRACE: Lifelong model editing with discrete key-value adaptors. arXiv preprint arXiv:2211.11031 (2023)

arXiv 2023
[9]

In: International Conference on Learning Representations (2022)

Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022)

2022
[10]

In: Proceedings of the 21st Conference on Computational Nat- ural Language Learning

Levy, O., Seo, M., Choi, E., Zettlemoyer, L.: Zero-shot relation extraction via read- ing comprehension. In: Proceedings of the 21st Conference on Computational Nat- ural Language Learning. pp. 333–342 (2017). https://doi.org/10.18653/v1/K17- 1034

work page doi:10.18653/v1/k17- 2017
[11]

In: International Conference on Learning Representations (2025) 14 Zhang et al

Ma, J.Y., Wang, H., Xu, H.X., Ling, Z.H., Gu, J.C.: Perturbation-restrained se- quential model editing. In: International Conference on Learning Representations (2025) 14 Zhang et al

2025
[12]

In: Advances in Neural Information Processing Systems

Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual asso- ciations in GPT. In: Advances in Neural Information Processing Systems. vol. 35 (2022)

2022
[13]

In: International Conference on Learning Representations (2023)

Meng, K., Sharma, A.S., Andonian, A., Belinkov, Y., Bau, D.: Mass-editing mem- ory in a transformer. In: International Conference on Learning Representations (2023)

2023
[14]

In: International Conference on Learning Representations (2022)

Mitchell, E., Lin, C., Bosselut, A., Manning, C.D., Finn, C.: Fast model editing at scale. In: International Conference on Learning Representations (2022)

2022
[15]

In: Proceedings of the 39th International Conference on Machine Learning

Mitchell, E., Lin, C., Bosselut, A., Manning, C.D., Finn, C.: Memory-based model editing at scale. In: Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 15817–15831 (2022)

2022
[16]

In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing

Reimers, N., Gurevych, I.: Sentence-BERT: Sentence embeddings using siamese BERT-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing. pp. 3982–3992 (2019). https://doi.org/10.18653/v1/D19- 1410

work page doi:10.18653/v1/d19- 2019
[17]

In: Advances in Neural Information Processing Systems (2025)

Seong, J., Park, J., Liermann, W., Choi, H., Nam, Y., Kim, H., Lim, S., Lee, N.: MemEIC: A step toward continual and compositional knowledge editing. In: Advances in Neural Information Processing Systems (2025)

2025
[18]

arXiv preprint arXiv:2311.04661 (2023)

Tan, C., Zhang, G., Fu, J.: Massive editing for large language models via meta learning. arXiv preprint arXiv:2311.04661 (2023)

arXiv 2023
[19]

In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (2024)

Wang, H., Liu, T., Li, R., Cheng, M.X., Zhao, T., Gao, J.: RoseLoRA: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowl- edge editing and fine-tuning. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (2024)

2024
[20]

In: Advances in Neural Information Processing Systems (2024)

Wang, P., Li, Z., Zhang, N., Xu, Z., Yao, Y., Jiang, Y., Xie, P., Huang, F., Chen, H.: WISE: Rethinking the knowledge memory for lifelong model editing of large language models. In: Advances in Neural Information Processing Systems (2024)

2024
[21]

E asy E dit: An Easy-to-use Knowledge Editing Framework for Large Language Models

Wang, P., Zhang, N., Tian, B., Xi, Z., Yao, Y., Xu, Z., Wang, M., Mao, S., Wang, X., Cheng, S., Liu, K., Ni, Y., Zheng, G., Chen, H.: EasyEdit: An easy-to-use knowledge editing framework for large language models. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (System Demonstrations). pp. 82–93 (2024). https://...

work page doi:10.18653/v1/2024.acl-demos.9 2024
[22]

In: Advances in Neural Information Processing Systems

Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: MiniLM: Deep self- attention distillation for task-agnostic compression of pre-trained transformers. In: Advances in Neural Information Processing Systems. vol. 33 (2020)

2020
[23]

arXiv preprint arXiv:2309.07597 (2023)

Xiao, S., Liu, Z., Zhang, P., Muennighoff, N.: C-Pack: Packaged resources to ad- vance general chinese embedding. arXiv preprint arXiv:2309.07597 (2023)

Pith/arXiv arXiv 2023
[24]

arXiv preprint arXiv:2406.11566 (2024)

Xie, J., Cao, P., Chen, Y., Chen, Y., Liu, K., Zhao, J.: MEMLA: Enhancing multi- lingual knowledge editing with neuron-masked low-rank adaptation. arXiv preprint arXiv:2406.11566 (2024)

arXiv 2024
[25]

arXiv preprint arXiv:2505.09388 (2025)

Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

Pith/arXiv arXiv 2025
[26]

In: International Conference on Learning Representations (2026)

Yang, W., Tang, R., Zang, H., Su, D., Cao, Q., Wang, J., Shen, H., Cheng, X., Sun, F.: Fine-tuning done right in model editing. In: International Conference on Learning Representations (2026)

2026
[27]

Zheng, C., Li, L., Dong, Q., Fan, Y., Wu, Z., Xu, J., Chang, B.: Can we edit factual knowledge by in-context learning? In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 4862–4876 (2023). https://doi.org/10.18653/v1/2023.emnlp-main.296 Route-Specialized Dual Adapters 15

work page doi:10.18653/v1/2023.emnlp-main.296 2023
[28]

MQ u AKE : Assessing Knowledge Editing in Language Models via Multi-Hop Questions

Zhong, Z., Wu, Z., Manning, C.D., Potts, C., Chen, D.: MQuAKE: Assessing knowledge editing in language models via multi-hop questions. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 15686–15702 (2023). https://doi.org/10.18653/v1/2023.emnlp-main.971

work page doi:10.18653/v1/2023.emnlp-main.971 2023
[29]

arXiv preprint arXiv:2012.00363 (2020)

Zhu, C., Rawat, A.S., Zaheer, M., Bhojanapalli, S., Li, D., Yu, F.X., Kumar, S.: Modifying memories in transformer models. arXiv preprint arXiv:2012.00363 (2020)

arXiv 2012

[1] [1]

Transac- tions of the Association for Computational Linguistics12, 283–298 (2024)

Cohen, R., Biran, E., Yoran, O., Globerson, A., Geva, M.: Evaluat- ing the ripple effects of knowledge editing in language models. Transac- tions of the Association for Computational Linguistics12, 283–298 (2024). https://doi.org/10.1162/tacl_a_00644 Route-Specialized Dual Adapters 13 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Trainable parameters (millions) 0....

work page doi:10.1162/tacl_a_00644 2024

[2] [2]

In: Proceedings of the 2021 Confer- ence on Empirical Methods in Natural Language Processing

De Cao, N., Aziz, W., Titov, I.: Editing factual knowledge in language models. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Lan- guage Processing. pp. 6491–6506 (2021). https://doi.org/10.18653/v1/2021.emnlp- main.522

work page doi:10.18653/v1/2021.emnlp- 2021

[3] [3]

In: Advances in Neural Information Processing Systems

Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: QLoRA: Efficient fine- tuning of quantized LLMs. In: Advances in Neural Information Processing Systems. vol. 36 (2023)

2023

[4] [4]

arXiv preprint arXiv:2407.21783 (2024)

Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The Llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

Pith/arXiv arXiv 2024

[5] [5]

In: In- ternational Conference on Learning Representations (2025)

Fang, J., Jiang, H., Wang, K., Ma, Y., Shi, J., Wang, X., He, X., Chua, T.S.: AlphaEdit: Null-space constrained knowledge editing for language models. In: In- ternational Conference on Learning Representations (2025)

2025

[6] [6]

arXiv preprint arXiv:2406.04496 (2024)

Ge, X., Mousavi, A., Grave, E., Joulin, A., Qian, K., Han, B., Arefiyan, M., Li, Y.: Time sensitive knowledge editing through efficient finetuning. arXiv preprint arXiv:2406.04496 (2024)

arXiv 2024

[7] [7]

arXiv preprint arXiv:2502.19416 (2025)

Gupta, A., Fang, C., Ozdemir, A., Lu, M., Alaa, A., Hartvigsen, T., Anu- manchipalli, G.: Norm growth and stability challenges in localized sequential knowledge editing. arXiv preprint arXiv:2502.19416 (2025)

arXiv 2025

[8] [8]

arXiv preprint arXiv:2211.11031 (2023)

Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., Ghassemi, M.: Ag- ing with GRACE: Lifelong model editing with discrete key-value adaptors. arXiv preprint arXiv:2211.11031 (2023)

arXiv 2023

[9] [9]

In: International Conference on Learning Representations (2022)

Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022)

2022

[10] [10]

In: Proceedings of the 21st Conference on Computational Nat- ural Language Learning

Levy, O., Seo, M., Choi, E., Zettlemoyer, L.: Zero-shot relation extraction via read- ing comprehension. In: Proceedings of the 21st Conference on Computational Nat- ural Language Learning. pp. 333–342 (2017). https://doi.org/10.18653/v1/K17- 1034

work page doi:10.18653/v1/k17- 2017

[11] [11]

In: International Conference on Learning Representations (2025) 14 Zhang et al

Ma, J.Y., Wang, H., Xu, H.X., Ling, Z.H., Gu, J.C.: Perturbation-restrained se- quential model editing. In: International Conference on Learning Representations (2025) 14 Zhang et al

2025

[12] [12]

In: Advances in Neural Information Processing Systems

Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual asso- ciations in GPT. In: Advances in Neural Information Processing Systems. vol. 35 (2022)

2022

[13] [13]

In: International Conference on Learning Representations (2023)

Meng, K., Sharma, A.S., Andonian, A., Belinkov, Y., Bau, D.: Mass-editing mem- ory in a transformer. In: International Conference on Learning Representations (2023)

2023

[14] [14]

In: International Conference on Learning Representations (2022)

Mitchell, E., Lin, C., Bosselut, A., Manning, C.D., Finn, C.: Fast model editing at scale. In: International Conference on Learning Representations (2022)

2022

[15] [15]

In: Proceedings of the 39th International Conference on Machine Learning

Mitchell, E., Lin, C., Bosselut, A., Manning, C.D., Finn, C.: Memory-based model editing at scale. In: Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 15817–15831 (2022)

2022

[16] [16]

In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing

Reimers, N., Gurevych, I.: Sentence-BERT: Sentence embeddings using siamese BERT-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing. pp. 3982–3992 (2019). https://doi.org/10.18653/v1/D19- 1410

work page doi:10.18653/v1/d19- 2019

[17] [17]

In: Advances in Neural Information Processing Systems (2025)

Seong, J., Park, J., Liermann, W., Choi, H., Nam, Y., Kim, H., Lim, S., Lee, N.: MemEIC: A step toward continual and compositional knowledge editing. In: Advances in Neural Information Processing Systems (2025)

2025

[18] [18]

arXiv preprint arXiv:2311.04661 (2023)

Tan, C., Zhang, G., Fu, J.: Massive editing for large language models via meta learning. arXiv preprint arXiv:2311.04661 (2023)

arXiv 2023

[19] [19]

In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (2024)

Wang, H., Liu, T., Li, R., Cheng, M.X., Zhao, T., Gao, J.: RoseLoRA: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowl- edge editing and fine-tuning. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (2024)

2024

[20] [20]

In: Advances in Neural Information Processing Systems (2024)

Wang, P., Li, Z., Zhang, N., Xu, Z., Yao, Y., Jiang, Y., Xie, P., Huang, F., Chen, H.: WISE: Rethinking the knowledge memory for lifelong model editing of large language models. In: Advances in Neural Information Processing Systems (2024)

2024

[21] [21]

E asy E dit: An Easy-to-use Knowledge Editing Framework for Large Language Models

Wang, P., Zhang, N., Tian, B., Xi, Z., Yao, Y., Xu, Z., Wang, M., Mao, S., Wang, X., Cheng, S., Liu, K., Ni, Y., Zheng, G., Chen, H.: EasyEdit: An easy-to-use knowledge editing framework for large language models. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (System Demonstrations). pp. 82–93 (2024). https://...

work page doi:10.18653/v1/2024.acl-demos.9 2024

[22] [22]

In: Advances in Neural Information Processing Systems

Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: MiniLM: Deep self- attention distillation for task-agnostic compression of pre-trained transformers. In: Advances in Neural Information Processing Systems. vol. 33 (2020)

2020

[23] [23]

arXiv preprint arXiv:2309.07597 (2023)

Xiao, S., Liu, Z., Zhang, P., Muennighoff, N.: C-Pack: Packaged resources to ad- vance general chinese embedding. arXiv preprint arXiv:2309.07597 (2023)

Pith/arXiv arXiv 2023

[24] [24]

arXiv preprint arXiv:2406.11566 (2024)

Xie, J., Cao, P., Chen, Y., Chen, Y., Liu, K., Zhao, J.: MEMLA: Enhancing multi- lingual knowledge editing with neuron-masked low-rank adaptation. arXiv preprint arXiv:2406.11566 (2024)

arXiv 2024

[25] [25]

arXiv preprint arXiv:2505.09388 (2025)

Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025)

Pith/arXiv arXiv 2025

[26] [26]

In: International Conference on Learning Representations (2026)

Yang, W., Tang, R., Zang, H., Su, D., Cao, Q., Wang, J., Shen, H., Cheng, X., Sun, F.: Fine-tuning done right in model editing. In: International Conference on Learning Representations (2026)

2026

[27] [27]

Zheng, C., Li, L., Dong, Q., Fan, Y., Wu, Z., Xu, J., Chang, B.: Can we edit factual knowledge by in-context learning? In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 4862–4876 (2023). https://doi.org/10.18653/v1/2023.emnlp-main.296 Route-Specialized Dual Adapters 15

work page doi:10.18653/v1/2023.emnlp-main.296 2023

[28] [28]

MQ u AKE : Assessing Knowledge Editing in Language Models via Multi-Hop Questions

Zhong, Z., Wu, Z., Manning, C.D., Potts, C., Chen, D.: MQuAKE: Assessing knowledge editing in language models via multi-hop questions. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 15686–15702 (2023). https://doi.org/10.18653/v1/2023.emnlp-main.971

work page doi:10.18653/v1/2023.emnlp-main.971 2023

[29] [29]

arXiv preprint arXiv:2012.00363 (2020)

Zhu, C., Rawat, A.S., Zaheer, M., Bhojanapalli, S., Li, D., Yu, F.X., Kumar, S.: Modifying memories in transformer models. arXiv preprint arXiv:2012.00363 (2020)

arXiv 2012