pith. sign in

arxiv: 2607.01978 · v1 · pith:NIC2OWTCnew · submitted 2026-07-02 · 💻 cs.AI · cs.CL· cs.CV

Multimodal Knowledge Edit-Scoped Generalization for Online Recursive MLLM Editing

Siyuan Li , Youyuan Zhang , Ruitong Liu , Junxi Wang , Jing Li This is my paper

Pith reviewed 2026-07-03 13:50 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.CV
keywords multimodal knowledge editingonline recursive editingMLLMedit scopecross-modal generalizationlocalitylow-rank updates
0
0 comments X

The pith

ScopeEdit decomposes each MLLM edit into a local absorption branch and an evidence-gated generalization branch to control semantic boundaries.

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

The paper establishes that reliable instance-level edits in multimodal large language models often fail to transfer correctly across modalities or stay contained, because they lack explicit control over propagation scope. It reframes the task as Edit-Scoped Generalization and introduces ScopeEdit to enforce that control by splitting each update into a modality-local branch for stable absorption and a shared branch that propagates only when visual and textual evidence align. Both branches use orthogonal low-rank write geometries and recursive Sherman-Morrison preconditioners to keep per-edit cost constant. Experiments on diverse benchmarks and long edit streams show the method improves the transfer-locality trade-off without harming reliability or efficiency. A reader would care because continual correction of multimodal models is practical only if each change stays inside its intended semantic region.

Core claim

The paper claims that reframing online MLLM editing as explicit control of each edit's propagation boundary, achieved by ScopeEdit's decomposition into a modality-local absorption branch and an evidence-gated shared generalization branch that operate in orthogonal low-rank spaces with Sherman-Morrison recursions, produces consistent gains in in-scope cross-modal transfer while preserving out-of-scope locality, edit reliability, stability, and constant overhead across benchmarks, backbones, and real-world VLKEB streams.

What carries the argument

ScopeEdit, the scope-aware editor that splits every update into a modality-local absorption branch and an evidence-gated shared generalization branch performing scope-separated writes in orthogonal low-rank spaces maintained by Sherman-Morrison recursions.

If this is right

  • Each edit can be absorbed locally while cross-modal variants receive the update only when evidence alignment passes the gate.
  • Out-of-scope leakage remains low even after hundreds of sequential edits because the two branches write in separate low-rank spaces.
  • Per-edit runtime and memory stay constant because preconditioners are updated recursively rather than recomputed.
  • The same decomposition applies across different MLLM architectures and to real-world vision-language knowledge editing benchmarks.
  • Edit reliability and long-horizon stability are retained while the scope trade-off improves.

Where Pith is reading between the lines

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

  • The same branch separation could be tested on non-multimodal continual editing tasks if a suitable evidence-alignment signal is defined for text-only or image-only streams.
  • If deeper semantic layers consistently host edit-related activity, future editors might locate the write space directly from activation maps rather than relying on fixed low-rank choices.
  • The evidence gate offers a concrete mechanism that could be combined with retrieval-augmented editing pipelines to decide when external context should influence scope.
  • Long-horizon stability under ScopeEdit suggests the method may reduce the frequency of full model resets in deployed multimodal systems that receive ongoing corrections.

Load-bearing premise

That alignment of visual and textual evidence can be measured reliably enough to gate cross-modal propagation without creating new instabilities or missing valid transfers.

What would settle it

An experiment on the same long-horizon edit streams and MLLM backbones in which ScopeEdit produces no measurable improvement in the in-scope transfer versus out-of-scope leakage trade-off compared with prior editors that lack the gated branch.

Figures

Figures reproduced from arXiv: 2607.01978 by Jing Li, Junxi Wang, Ruitong Liu, Siyuan Li, Youyuan Zhang.

Figure 1
Figure 1. Figure 1: Motivation and formulation of edit-scoped generalization in online MLLM editing. (A) Online multimodal knowledge editing incorporates a stream of visual-textual corrections into an MLLM. (B) Existing editors mainly focus on edit reliability and long-horizon stability, but lack explicit control over the semantic scope of each edit, leading to under-generalization on in-scope variants or over-generalization … view at source ↗
Figure 2
Figure 2. Figure 2: Pilot analysis of edit-scoped generalization and its layer-wise emergence in online MLLM editing, using M-ORE on LLaVA-v1.5. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity analysis of the low-rank write-interface dimension [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sequential visualization of hidden-state distribution shifts under long-horizon online editing. We compare MEND, M-ORE, and ScopeEdit on locality [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pairwise weight-entanglement analysis across consecutive online edits. The first row shows results on BLIP-2, and the second row shows results [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Efficiency metrics of editors on LLaVA-v1.5 over the online edit sequence [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-model validation of the layer-wise phenomenon. BLIP2-OPT, Qwen2-VL, and Qwen3-VL exhibit similar late-layer concentration patterns in [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Case study of edit-scope control. Given the same multimodal edit, ScopeEdit correctly absorbs the target correction and generalizes it to in-scope text [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
read the original abstract

Online multimodal knowledge editing requires injecting a continual stream of visual-textual corrections into multimodal large language models (MLLMs) with bounded overhead and minimal disruption to unrelated behaviors. Existing editors mainly emphasize edit reliability and long-horizon stability, but rarely control the semantic boundary of each edit. Our pilot analyses of post-edit behaviors and internal neuronal activities reveal a scope gap behind reliable edits: instance-level success neither guarantees transfer to valid cross-modal variants nor prevents leakage to unrelated inputs, while edit-related cross-modal responses concentrate in deeper semantic layers. Therefore, we formulate Edit-Scoped Generalization, reframing online MLLM editing from merely correcting an instance to controlling the propagation boundary of each edit. To this end, we propose ScopeEdit, a scope-aware online editor that decomposes each update into a modality-local absorption branch and an evidence-gated shared generalization branch. The local branch supports stable edit absorption, whereas the shared branch enables cross-modal propagation only when visual and textual evidence are sufficiently aligned. Both branches perform scope-separated write geometries in orthogonal low-rank spaces and maintain branch-wise preconditioners via Sherman--Morrison recursions, yielding constant per-edit overhead. Extensive experiments across diverse benchmarks, long-horizon edit streams, MLLM backbones, real-world VLKEB scenarios, and complex vision-language architectures show that ScopeEdit consistently improves the trade-off between in-scope cross-modal transfer and out-of-scope locality, while preserving edit reliability, stability and online efficiency. Our code is available at https://github.com/lab-klc/ScopeEdit.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 3 minor

Summary. The paper identifies a 'scope gap' in existing online MLLM editors where instance-level edits succeed but fail to generalize appropriately to cross-modal variants or leak to unrelated inputs. It formulates Edit-Scoped Generalization and proposes ScopeEdit, which decomposes each update into a modality-local absorption branch and an evidence-gated shared generalization branch. Both branches use orthogonal low-rank updates maintained via Sherman-Morrison recursions for constant per-edit overhead. Extensive experiments across benchmarks, long-horizon streams, multiple MLLM backbones, and VLKEB scenarios are claimed to show improved in-scope cross-modal transfer versus out-of-scope locality while preserving reliability, stability, and efficiency. Code is released.

Significance. If the empirical results hold, the work addresses a practically important limitation in multimodal knowledge editing by explicitly controlling edit propagation boundaries rather than only optimizing reliability and stability. The decomposition into local and gated shared branches with orthogonal low-rank geometry and recursive preconditioners is internally consistent with the constant-overhead goal. Credit is due for the public code release and the breadth of reported experiments across architectures and real-world scenarios.

minor comments (3)
  1. The abstract states that edit-related responses concentrate in deeper semantic layers, but the precise layer indices or activation statistics used to motivate the shared-branch design are not referenced; adding a pointer to the relevant figure or table would strengthen the motivation.
  2. Notation for the evidence-alignment gate (e.g., the threshold or similarity measure deciding when the shared branch activates) should be introduced with an equation number in the method section for reproducibility.
  3. Table captions or legends should explicitly define the 'in-scope transfer' and 'out-of-scope locality' metrics used to quantify the claimed trade-off improvement.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work on addressing the scope gap in online MLLM editing via Edit-Scoped Generalization and ScopeEdit. The recommendation for minor revision is noted, and we appreciate the recognition of the decomposition into modality-local and evidence-gated branches with orthogonal low-rank updates and recursive maintenance for efficiency.

Circularity Check

0 steps flagged

No significant circularity; empirical engineering contribution without reduction to inputs

full rationale

The paper introduces ScopeEdit as a scope-aware editor via explicit design choices: decomposition into modality-local absorption and evidence-gated shared generalization branches, orthogonal low-rank spaces, and Sherman-Morrison recursions for preconditioners. These are presented as constructions to achieve constant overhead and scope separation, not as derivations from prior fitted quantities or self-citations. No equations or steps reduce by construction to the target claims (e.g., no fitted parameters renamed as predictions, no uniqueness theorems imported from self-citations). Central claims rest on experimental results across benchmarks, architectures, and scenarios, which are independent of any internal reduction. This matches the default expectation of a non-circular empirical method.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the unverified pilot finding of a scope gap and the engineering assumption that the proposed branch decomposition plus evidence gating will enforce the desired propagation boundary; no free parameters or invented physical entities are mentioned.

axioms (1)
  • domain assumption Edit-related cross-modal responses concentrate in deeper semantic layers
    Invoked as the basis for the shared generalization branch design.
invented entities (1)
  • ScopeEdit no independent evidence
    purpose: Scope-aware online editor using modality-local and evidence-gated branches
    New method introduced by the paper; no independent evidence outside the claimed experiments.

pith-pipeline@v0.9.1-grok · 5822 in / 1377 out tokens · 41648 ms · 2026-07-03T13:50:52.494412+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

62 extracted references · 8 canonical work pages · 5 internal anchors

  1. [1]

    LLaMA: Open and Efficient Foundation Language Models

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    A Survey of Large Language Models

    W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y . Hou, Y . Min, B. Zhang, J. Zhang, Z. Donget al., “A survey of large language models,”arXiv preprint arXiv:2303.18223, vol. 1, no. 2, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” inProceedings of the International conference on machine learning (ICML), 2023, pp. 19 730–19 742

    2023

  4. [4]

    Video-llava: Learning united visual representation by alignment before projection,

    B. Lin, Y . Ye, B. Zhu, J. Cui, M. Ning, P. Jin, and L. Yuan, “Video-llava: Learning united visual representation by alignment before projection,” inProceedings of the 2024 conference on empirical methods in natural language processing (EMNLP), 2024, pp. 5971–5984

    2024

  5. [5]

    Moe-llava: Mixture of experts for large vision- language models,

    B. Lin, Z. Tang, Y . Ye, J. Huang, J. Zhang, Y . Pang, P. Jin, M. Ning, J. Luo, and L. Yuan, “Moe-llava: Mixture of experts for large vision- language models,”IEEE Transactions on Multimedia, 2026

    2026

  6. [6]

    Vision-controllable language model for image-guided story ending generation,

    D. Xue, S. Qian, and C. Xu, “Vision-controllable language model for image-guided story ending generation,”IEEE Transactions on Multime- dia, 2026

    2026

  7. [7]

    When large multimodal models confront evolving knowledge: Challenges and explorations,

    K. Jiang, Y . Du, Y . Ding, Y . Ren, N. Jiang, Z. Gao, Z. Zheng, L. Liu, B. Li, and Q. Li, “When large multimodal models confront evolving knowledge: Challenges and explorations,” inProceedings of the International Conference on Learning Representations (ICLR), 2026

    2026

  8. [8]

    Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model,

    S. Wang, X. Cao, J. Zhang, Z. Yuan, S. Shan, X. Chen, and W. Gao, “Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  9. [9]

    Locating and editing factual associations in gpt,

    K. Meng, D. Bau, A. Andonian, and Y . Belinkov, “Locating and editing factual associations in gpt,” inProceedings of the Advances in neural information processing systems (NeurIPS), 2022, pp. 17 359–17 372

    2022

  10. [10]

    Alphaedit: Null-space constrained knowledge editing for language models,

    J. Fang, H. Jiang, K. Wang, Y . Ma, J. Shi, X. Wang, X. He, and T.-S. Chua, “Alphaedit: Null-space constrained knowledge editing for language models,” inProceedings of the International Conference on Learning Representations (ICLR), 2025

    2025

  11. [11]

    Attribution analysis meets model editing: Advancing knowledge correction in vision language models with visedit,

    Q. Chen, T. Zhang, C. Wang, X. He, D. Wang, and T. Liu, “Attribution analysis meets model editing: Advancing knowledge correction in vision language models with visedit,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025, pp. 2168–2176

    2025

  12. [12]

    Lifelong knowledge editing for vision language models with low-rank mixture-of- experts,

    Q. Chen, C. Wang, D. Wang, T. Zhang, W. Li, and X. He, “Lifelong knowledge editing for vision language models with low-rank mixture-of- experts,” inProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025, pp. 9455–9466

    2025

  13. [13]

    Can we edit multimodal large language models?

    S. Cheng, B. Tian, Q. Liu, X. Chen, Y . Wang, H. Chen, and N. Zhang, “Can we edit multimodal large language models?” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023, pp. 13 877–13 888

    2023

  14. [14]

    Principled multimodal representation learning,

    X. Liu, X. Xia, S.-K. Ng, and T.-S. Chua, “Principled multimodal representation learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  15. [15]

    Modality-Decoupled Online Recursive Editing

    S. Li, Y . Zhang, F. Liu, and J. Li, “Modality-decoupled online recursive editing,”arXiv preprint arXiv:2605.20273, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Dualedit: Dual editing for knowledge updating in vision-language models,

    Z. Shi, B. Wang, C. Si, Y . Wu, J. Kim, and H. Pfister, “Dualedit: Dual editing for knowledge updating in vision-language models,” in Proceedings of the Second Conference on Language Modeling (COLM), 2025

    2025

  17. [17]

    Can we continually edit language models? on the knowledge attenuation in sequential model editing,

    Q. Li and X. Chu, “Can we continually edit language models? on the knowledge attenuation in sequential model editing,” inFindings of the Association for Computational Linguistics (ACL), 2024, pp. 5438–5455

    2024

  18. [18]

    VLKEB: A large vision-language model knowledge editing bench- mark,

    H. Huang, H. Zhong, T. Yu, Q. Liu, S. Wu, L. Wang, and T. Tan, “VLKEB: A large vision-language model knowledge editing bench- mark,” inAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024

    2024

  19. [19]

    Can knowledge be transferred from unimodal to multi- modal? investigating the transitivity of multimodal knowledge editing,

    L. Fang, X. Wang, D. Wang, Z. Wu, Y . Guo, H. Zhu, Z. Zhang, and G. Liu, “Can knowledge be transferred from unimodal to multi- modal? investigating the transitivity of multimodal knowledge editing,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 2482–2490

    2025

  20. [20]

    Mmke- bench: A multimodal editing benchmark for diverse visual knowledge,

    Y . Du, K. Jiang, Z. Gao, C. Shi, Z. Zheng, S. Qi, and Q. Li, “Mmke- bench: A multimodal editing benchmark for diverse visual knowledge,” inThe Thirteenth International Conference on Learning Representa- tions(ICLR), 2025

    2025

  21. [21]

    MC-MKE: A fine-grained multimodal knowledge editing benchmark emphasizing modality consistency,

    J. Zhang, H. Zhang, X. Yin, B. Huang, X. Zhang, X. Hu, and X. Wan, “MC-MKE: A fine-grained multimodal knowledge editing benchmark emphasizing modality consistency,” inFindings of the Association for Computational Linguistics(ACL), 2025, pp. 17 430–17 445

    2025

  22. [22]

    Memeic: A step toward continual and compositional knowl- edge editing,

    J. Seong, J. Park, W. Liermann, H. Choi, Y . Nam, H. Kim, S. Lim, and N. Lee, “Memeic: A step toward continual and compositional knowl- edge editing,” inProceedings of the Advances in neural information processing systems (NeurIPS), vol. 38, 2025, pp. 129 205–129 242

    2025

  23. [23]

    Exploring and evaluat- ing multimodal knowledge reasoning consistency of multimodal large language models,

    B. Jia, J. Zhang, H. Zhang, and X. Wan, “Exploring and evaluat- ing multimodal knowledge reasoning consistency of multimodal large language models,” inFindings of the Association for Computational Linguistics(EMNLP), 2025, pp. 11 966–11 981

    2025

  24. [24]

    M2Edit: Locate and edit multi-granularity knowledge in multimodal large language model,

    Y . Zhou, P. Cao, Y . Chen, Q. Liu, D. Sui, X. Chen, K. Liu, and J. Zhao, “M2Edit: Locate and edit multi-granularity knowledge in multimodal large language model,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing(EMNLP), 2025, pp. 29 029–29 042

    2025

  25. [25]

    V-SEAM: Visual semantic editing and attention modulating for causal interpretability of vision-language models,

    Q. Wang, J. Hu, and M. Jiang, “V-SEAM: Visual semantic editing and attention modulating for causal interpretability of vision-language models,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing(EMNLP), 2025, pp. 17 396–17 420

    2025

  26. [26]

    Dsca: Dynamic subspace concept alignment for lifelong vlm editing,

    G. Das and S. Jena, “Dsca: Dynamic subspace concept alignment for lifelong vlm editing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 40 772– 40 781

    2026

  27. [27]

    Insight over sight: Exploring the vision-knowledge conflicts in multimodal LLMs,

    X. Liu, W. Wang, Y . Yuan, J.-t. Huang, Q. Liu, P. He, and Z. Tu, “Insight over sight: Exploring the vision-knowledge conflicts in multimodal LLMs,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025, pp. 17 825–17 846

    2025

  28. [28]

    Is cognition consistent with perception? assessing and mitigating multimodal knowledge conflicts in document understanding,

    Z. Shao, F. Gao, Z. Zhu, C. Luo, H. Xing, Z. Yu, Q. Zheng, M. Yan, and J. Bu, “Is cognition consistent with perception? assessing and mitigating multimodal knowledge conflicts in document understanding,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing(EMNLP), 2025, pp. 30 923–30 944

    2025

  29. [29]

    Knowledge neurons in pretrained transformers,

    D. Dai, L. Dong, Y . Hao, Z. Sui, B. Chang, and F. Wei, “Knowledge neurons in pretrained transformers,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), 2022, pp. 8493–8502

    2022

  30. [30]

    Editable neural networks,

    A. Sinitsin, V . Plokhotnyuk, D. V . Pyrkin, S. Popov, and A. Babenko, “Editable neural networks,” inProceedings of the International Confer- ence on Learning Representations (ICLR), 2020

    2020

  31. [31]

    Zhang, Y

    N. Zhang, Y . Yao, B. Tian, P. Wang, S. Deng, M. Wang, Z. Xi, S. Mao, J. Zhang, Y . Niet al., “A comprehensive study of knowledge editing for large language models,”arXiv preprint arXiv:2401.01286, 2024

    work page arXiv 2024

  32. [32]

    Transformer-patcher: One mistake worth one neuron,

    Z. Huang, Y . Shen, X. Zhang, J. Zhou, W. Rong, and Z. Xiong, “Transformer-patcher: One mistake worth one neuron,” inProceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  33. [33]

    Neighboring perturba- tions of knowledge editing on large language models,

    J.-Y . Ma, Z.-H. Ling, N. Zhang, and J.-C. Gu, “Neighboring perturba- tions of knowledge editing on large language models,” inProceedings of the 41st International Conference on Machine Learning(ICML), vol. 235, 2024, pp. 33 839–33 854

    2024

  34. [34]

    Uncovering over- fitting in large language model editing,

    M. Zhang, X. Ye, Q. Liu, S. Wu, P. Ren, and Z. Chen, “Uncovering over- fitting in large language model editing,” inProceedings of the Thirteenth International Conference on Learning Representations(ICLR), 2025. 16

    2025

  35. [35]

    Revealing and mitigating over-attention in knowledge editing,

    P. Wang, Z. Tang, K. Zhou, J. Li, Q. Zhu, and M. Zhang, “Revealing and mitigating over-attention in knowledge editing,” inProceedings of the Thirteenth International Conference on Learning Representa- tions(ICLR), 2025, pp. 20 922–20 948

    2025

  36. [36]

    Parameter-aware contrastive knowledge editing: Tracing and rectifying based on critical transmission paths,

    S. Zhai, Y . Meng, Y . Zhang, and G. Qi, “Parameter-aware contrastive knowledge editing: Tracing and rectifying based on critical transmission paths,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025, pp. 28 189–28 200

    2025

  37. [37]

    CaKE: Circuit-aware editing enables generalizable knowledge learn- ers,

    Y . Yao, J. Fang, J.-C. Gu, N. Zhang, S. Deng, H. Chen, and N. Peng, “CaKE: Circuit-aware editing enables generalizable knowledge learn- ers,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing(EMNLP), 2025, pp. 11 366–11 382

    2025

  38. [38]

    Mass-editing memory in a transformer,

    K. Meng, A. S. Sharma, A. J. Andonian, Y . Belinkov, and D. Bau, “Mass-editing memory in a transformer,” inProceedings of the Interna- tional Conference on Learning Representations (ICLR), 2023

    2023

  39. [39]

    Deltaedit: Enhancing sequential editing in large language models by controlling superimposed noise,

    D. Cao, Y . Cai, R. Guo, X. He, and G. Liu, “Deltaedit: Enhancing sequential editing in large language models by controlling superimposed noise,”arXiv preprint arXiv:2505.07899, 2025

    work page arXiv 2025

  40. [40]

    Editing factual knowledge in language models,

    N. De Cao, W. Aziz, and I. Titov, “Editing factual knowledge in language models,” inProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021

    2021

  41. [41]

    Fast model editing at scale,

    E. Mitchell, C. Lin, A. Bosselut, C. Finn, and C. D. Manning, “Fast model editing at scale,” inProceedings of the International Conference on Learning Representations (ICLR), 2022

    2022

  42. [42]

    SAKE: Steering activations for knowledge editing,

    M. Scialanga, T. Laugel, V . Grari, and M. Detyniecki, “SAKE: Steering activations for knowledge editing,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025, pp. 15 966–15 978

    2025

  43. [43]

    Knowledge editing through chain-of-thought,

    C. Wang, W. Su, Q. Ai, Y . Tang, and Y . Liu, “Knowledge editing through chain-of-thought,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing(EMNLP), 2025, pp. 10 673– 10 693

    2025

  44. [44]

    Can we edit factual knowledge by in-context learning?

    C. Zheng, L. Li, Q. Dong, Y . Fan, Z. Wu, J. Xu, and B. Chang, “Can we edit factual knowledge by in-context learning?” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023, pp. 4862–4876

    2023

  45. [45]

    Memory- based model editing at scale,

    E. Mitchell, C. Lin, A. Bosselut, C. D. Manning, and C. Finn, “Memory- based model editing at scale,” inProceedings of the International Conference on Machine Learning (ICML), 2022, pp. 15 817–15 831

    2022

  46. [46]

    Aging with GRACE: lifelong model editing with discrete key- value adaptors,

    T. Hartvigsen, S. Sankaranarayanan, H. Palangi, Y . Kim, and M. Ghas- semi, “Aging with GRACE: lifelong model editing with discrete key- value adaptors,” inProceedings of the Advances in neural information processing systems (NeurIPS), 2023

    2023

  47. [47]

    MELO: enhancing model editing with neuron-indexed dynamic lora,

    L. Yu, Q. Chen, J. Zhou, and L. He, “MELO: enhancing model editing with neuron-indexed dynamic lora,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2024, pp. 19 449–19 457

    2024

  48. [48]

    Llms meet multimodal generation and editing: A survey,

    Y . He, Z. Liu, J. Chen, Z. Tian, H. Liu, X. Chi, R. Liu, R. Yuan, Y . Xing, W. Wang, J. Dai, Y . Zhang, W. Xue, Q. Liu, Y . Guo, and Q. Chen, “Llms meet multimodal generation and editing: A survey,” arXiv preprint arXiv:2405.19334, 2024

    work page arXiv 2024

  49. [49]

    BalancEdit: Dynamically balancing the generality-locality trade-off in multi-modal model editing,

    D. Guo, M. Hu, Z. Guan, T. Hartvigsen, and S. Li, “BalancEdit: Dynamically balancing the generality-locality trade-off in multi-modal model editing,” inProceedings of the 42nd International Conference on Machine Learning(ICML), vol. 267, 2025, pp. 20 843–20 857

    2025

  50. [50]

    Visual- oriented fine-grained knowledge editing for multimodal large language models,

    Z. Zeng, L. Gu, X. Yang, Z. Duan, Z. Shi, and M. Wang, “Visual- oriented fine-grained knowledge editing for multimodal large language models,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 2491–2500

    2025

  51. [51]

    Unlocking efficient, scalable, and continual knowledge editing with basis-level representation fine-tuning,

    T. Liu, R. Li, Y . Qi, H. Liu, X. Tang, T. Zheng, Q. Yin, M. Cheng, J. Huan, H. Wang, and J. Gao, “Unlocking efficient, scalable, and continual knowledge editing with basis-level representation fine-tuning,” inProceedings of the Thirteenth International Conference on Learning Representations(ICLR), 2025, pp. 18 939–18 959

    2025

  52. [52]

    Neuron-level sequential editing for large language models,

    H. Jiang, J. Fang, T. Zhang, B. Bi, A. Zhang, R. Wang, T. Liang, and X. Wang, “Neuron-level sequential editing for large language models,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025, pp. 16 678–16 702

    2025

  53. [53]

    MEMORYLLM: Towards self-updatable large language models,

    Y . Wang, Y . Gao, X. Chen, H. Jiang, S. Li, J. Yang, Q. Yin, Z. Li, X. Li, B. Yin, J. Shang, and J. Mcauley, “MEMORYLLM: Towards self-updatable large language models,” inProceedings of the 41st In- ternational Conference on Machine Learning(ICML), 2024, pp. 50 453– 50 466

    2024

  54. [54]

    AdaEdit: Advancing continuous knowledge editing for large language models,

    Q. Li and X. Chu, “AdaEdit: Advancing continuous knowledge editing for large language models,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025, pp. 4127– 4149

    2025

  55. [55]

    Lifelong knowledge editing for LLMs with retrieval-augmented con- tinuous prompt learning,

    Q. Chen, T. Zhang, X. He, D. Li, C. Wang, L. Huang, and H. Xue’, “Lifelong knowledge editing for LLMs with retrieval-augmented con- tinuous prompt learning,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024, pp. 13 565–13 580

    2024

  56. [56]

    WISE: rethinking the knowledge memory for lifelong model editing of large language models,

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

    2024

  57. [57]

    Should we really edit language models? on the evaluation of edited language models,

    Q. Li, X. Liu, Z. Tang, P. Dong, Z. Li, X. Pan, and X. Chu, “Should we really edit language models? on the evaluation of edited language models,” inProceedings of the 38th International Conference on Neural Information Processing Systems(NeurIPS), 2024

    2024

  58. [58]

    Wik- ibigedit: Understanding the limits of lifelong knowledge editing in llms,

    L. Thede, K. Roth, M. Bethge, Z. Akata, and T. Hartvigsen, “Wik- ibigedit: Understanding the limits of lifelong knowledge editing in llms,” inProceedings of the International Conference on Machine Learning (ICML), 2025

    2025

  59. [59]

    Transformer feed-forward layers are key-value memories,

    M. Geva, R. Schuster, J. Berant, and O. Levy, “Transformer feed-forward layers are key-value memories,” inProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021, pp. 5484–5495

    2021

  60. [60]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved baselines with visual instruction tuning,” inProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2024, pp. 26 286–26 296

    2024

  61. [61]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y . Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin, “Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,”arXiv preprint arXiv:2409.12191, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  62. [62]

    Qwen3-VL Technical Report

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y . Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y . Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang,...

    work page internal anchor Pith review Pith/arXiv arXiv 2025