pith. sign in

arxiv: 2606.03924 · v1 · pith:E7IG7JAQnew · submitted 2026-06-02 · 💻 cs.CL

Knowledge Editing in Masked Diffusion Language Models

Haewon Park , Yohan Jo This is my paper

Pith reviewed 2026-06-28 09:49 UTC · model grok-4.3

classification 💻 cs.CL
keywords knowledge editingmasked diffusion modelscausal tracinglocate-then-editautoregressive modelsfactual knowledgemodel editingiterative denoising
0
0 comments X

The pith

The same early-to-mid MLP at the last subject token is the best edit site for facts in both masked diffusion and autoregressive language models.

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

The paper checks whether locate-then-edit, which finds a fact's location via causal tracing and then changes weights there, still works when the model generates text by iterative denoising instead of left-to-right prediction. Causal tracing points to the identical location in both model families, and edits placed there succeed for single-token facts in both. For longer targets the diffusion models lose accuracy because each denoising step must pass through partially revealed states that the edit never saw. The authors add a step that optimizes the edit across those intermediate states and largely close the gap.

Core claim

Causal tracing identifies the same early-to-mid-layer MLP at the last subject token as the most effective edit site in both MDMs and ARMs. Single-token edits succeed equally in the two families, yet multi-token targets degrade in MDMs because generation requires the model to produce the target through partially unmasked intermediate states for which the edit was never optimized. Optimizing the edit directly over those states restores most of the lost performance.

What carries the argument

Causal tracing to locate the edit site, followed by locate-then-edit weight updates adapted to the iterative denoising trajectory of masked diffusion models.

If this is right

  • The same causal-tracing procedure can be used to locate edit sites across autoregressive and masked diffusion families.
  • Single-token factual updates transfer directly once the shared location is used.
  • Multi-token edits in diffusion models require an auxiliary optimization pass over the denoising trajectory.
  • The location of stored facts is more stable across generation paradigms than the generation process itself.

Where Pith is reading between the lines

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

  • Editing methods developed for next-token models may need trajectory-aware corrections whenever generation is iterative or bidirectional.
  • The shared location suggests that fact storage is largely independent of whether a model predicts tokens sequentially or denoises masks.
  • The same correction could be tested on other non-autoregressive generators that rely on intermediate partially observed states.

Load-bearing premise

The performance gap on multi-token targets is caused by the need to generate through partially unmasked states that were never seen during the edit, rather than by other differences between the model families.

What would settle it

If applying the extra optimization over intermediate denoising states still leaves a large multi-token performance gap between the MDMs and the matched ARMs, the claimed cause would be falsified.

Figures

Figures reproduced from arXiv: 2606.03924 by Haewon Park, Yohan Jo.

Figure 1
Figure 1. Figure 1: Causal tracing across ARM/MDM pairs. Top row: ARM baselines; bottom row: their MDM counterparts. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Token position ablation for editing. L0-3 L8-11 L16-19 L24-27 L28-31 0.0 0.5 1.0 Efficacy pre = 0.01 LLaDA-8B L0-3 L8-11 L16-19 L24-27 0.0 0.5 1.0 pre = 0.00 Dream-v0-7B L0-3 L8-11 L16-19 L24-27 L28-31 0.0 0.5 1.0 Generalization pre = 0.01 L0-3 L8-11 L16-19 L24-27 0.0 0.5 1.0 pre = 0.00 L0-3 L8-11 L16-19 L24-27 L28-31 edited layer range 0.0 0.1 0.2 Specificity pre = 0.18 L0-3 L8-11 L16-19 L24-27 edited lay… view at source ↗
Figure 3
Figure 3. Figure 3: Layer ablation for editing. in Appendix D. Token position. We first vary the token position at which the edit is applied—the first and last sub￾ject tokens, the first token after the subject, and the first answer position—and measure editing per￾formance ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Layer sweep for ARM editing. 0.0 0.2 0.4 0.6 0.8 1.0 LLaMA-3-8B 0.40 0.95 0.42 0.78 Efficacy 0.0 0.2 0.4 0.6 0.8 1.0 0.17 0.56 0.04 0.06 Generalization 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.26 0.25 0.29 0.27 pre = 0.24 Specificity 1st subj Last subj 1st after last 0.0 0.2 0.4 0.6 0.8 1.0 Qwen2.5-7B 0.50 0.73 0.93 0.93 1st subj Last subj 1st after last 0.0 0.2 0.4 0.6 0.8 1.0 0.32 0.45 0.08 0.10 1st sub… view at source ↗
Figure 5
Figure 5. Figure 5: Token position ablation for ARM editing. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
read the original abstract

Knowledge editing aims to update or correct factual knowledge in a language model. A widely used approach, locate-then-edit, does this in two steps: it first localizes a fact within the model, then edits the weights there. To date, such methods have been developed exclusively on autoregressive models (ARMs). Whether their underlying assumptions hold for masked diffusion models (MDMs), which model text bidirectionally and generate by iterative denoising rather than next-token prediction, remains an open question. We address it by transferring locate-then-edit to MDMs and comparing two MDMs (LLaDA, Dream) with two ARMs (LLaMA, Qwen) at matched scale. Our central finding has two parts. First, where an edit is applied transfers across paradigms: causal tracing highlights the same early-to-mid-layer MLP at the last subject token in both, and editing is most effective there. Second, this shared location does not guarantee a shared outcome. Single-token edits succeed in both, but as targets grow longer, editing degrades systematically in the MDMs but not the ARMs. The failure stems from how the edited fact is generated: producing a multi-token target requires passing through partially unmasked intermediate states for which the edit was never optimized. Guided by this diagnosis, we introduce a simple correction that optimizes the edit for these states, substantially restoring multi-token performance.

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

2 major / 2 minor

Summary. The paper investigates whether locate-then-edit knowledge editing, previously developed only for autoregressive models (ARMs), transfers to masked diffusion models (MDMs). Using causal tracing on matched-scale models (LLaDA/Dream vs. LLaMA/Qwen), it finds that the effective edit location (early-to-mid-layer MLP at the last subject token) is shared across paradigms. Single-token edits succeed in both, but multi-token targets degrade in MDMs because generation requires passing through partially unmasked intermediate states never seen during editing; a proposed correction optimizes the edit for these states and restores performance.

Significance. If the empirical results hold after addressing the noted gaps, the work would usefully extend locate-then-edit methods to the MDM paradigm and supply a concrete mechanistic account of why multi-token editing behaves differently under iterative denoising. The shared location finding and the practical correction are the most immediately actionable contributions.

major comments (2)
  1. [Abstract and experimental sections] Abstract and the experimental comparison (LLaDA/Dream vs. LLaMA/Qwen): the central diagnosis that multi-token degradation arises specifically because the edited fact must be generated through partially unmasked intermediate states is not isolated from other uncontrolled differences between the model families (bidirectional vs. unidirectional attention, pretraining objectives, data mixtures). No ablation holds architecture and training fixed while varying only the denoising schedule, so the hypothesized mechanism remains the least secured link in the argument.
  2. [Method / correction description] The introduced correction is reported to restore multi-token performance, yet the manuscript provides no direct test (e.g., an ablation that applies the correction only to the hypothesized states versus other interventions) showing that the improvement occurs via the claimed mechanism rather than by compensating for any of the other architectural mismatches.
minor comments (2)
  1. [Abstract] The abstract states clear empirical findings but supplies no quantitative numbers, error bars, dataset sizes, or ablation controls; readers must reach the full experimental sections to evaluate effect sizes.
  2. [Method] Notation for the locate-then-edit procedure and the new correction should be introduced with explicit equations or pseudocode to make the optimization over intermediate states fully reproducible from the text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need to better isolate the proposed mechanism. We respond point-by-point to the major comments and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract and experimental sections] Abstract and the experimental comparison (LLaDA/Dream vs. LLaMA/Qwen): the central diagnosis that multi-token degradation arises specifically because the edited fact must be generated through partially unmasked intermediate states is not isolated from other uncontrolled differences between the model families (bidirectional vs. unidirectional attention, pretraining objectives, data mixtures). No ablation holds architecture and training fixed while varying only the denoising schedule, so the hypothesized mechanism remains the least secured link in the argument.

    Authors: We agree that the comparison across model families leaves the denoising schedule confounded with attention directionality, pretraining objectives, and data. The shared early-to-mid MLP location is nevertheless observed despite these differences. Fully isolating the iterative denoising effect would require training otherwise identical models that differ only in generation procedure, which exceeds the scope of the present study. In the revision we will add an explicit limitations paragraph qualifying the mechanistic claim and noting the absence of such a controlled ablation. revision: partial

  2. Referee: [Method / correction description] The introduced correction is reported to restore multi-token performance, yet the manuscript provides no direct test (e.g., an ablation that applies the correction only to the hypothesized states versus other interventions) showing that the improvement occurs via the claimed mechanism rather than by compensating for any of the other architectural mismatches.

    Authors: The correction is derived directly from the partial-mask states that arise only under MDM generation; its benefit is confined to multi-token targets in the MDM setting. We acknowledge, however, that the manuscript lacks an explicit ablation pitting the correction against generic interventions that do not target those states. In revision we will expand the method description to clarify the motivation and, space permitting, add a supplementary comparison in the appendix. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical transfer and comparisons

full rationale

The paper transfers causal tracing and locate-then-edit from prior external work, performs cross-paradigm experiments on matched-scale models, and diagnoses multi-token degradation via observed generation process in MDMs. No equations, fitted parameters, or self-citations are shown to reduce the central claims (location transfer or state-specific degradation) to inputs by construction. The derivation chain is self-contained against external benchmarks and does not invoke any of the enumerated circular patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on the standard assumption from prior editing literature that causal tracing reliably localizes factual knowledge; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption Causal tracing identifies the causal site of factual knowledge storage in the model weights
    Invoked to locate the edit site in both model classes.

pith-pipeline@v0.9.1-grok · 5768 in / 1269 out tokens · 26025 ms · 2026-06-28T09:49:50.142092+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

60 extracted references · 8 canonical work pages

  1. [1]

    On Faithfulness and Factuality in Abstractive Summarization

    Maynez, Joshua and Narayan, Shashi and Bohnet, Bernd and McDonald, Ryan. On Faithfulness and Factuality in Abstractive Summarization. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653/v1/2020.acl-main.173

    work page doi:10.18653/v1/2020.acl-main.173 2020

  2. [4]

    2024 , eprint=

    A Comprehensive Study of Knowledge Editing for Large Language Models , author=. 2024 , eprint=

    2024

  3. [5]

    2025 , eprint=

    Scaling Laws for Forgetting during Finetuning with Pretraining Data Injection , author=. 2025 , eprint=

    2025

  4. [6]

    2025 , eprint=

    An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning , author=. 2025 , eprint=

    2025

  5. [8]

    2024 , eprint=

    Knowledge Editing for Large Language Models: A Survey , author=. 2024 , eprint=

    2024

  6. [9]

    Automated Knowledge Base Construction , year=

    KAMEL: Knowledge Analysis with Multitoken Entities in Language Models , author=. Automated Knowledge Base Construction , year=

  7. [10]

    2023 , eprint=

    Locating and Editing Factual Associations in GPT , author=. 2023 , eprint=

    2023

  8. [11]

    The Eleventh International Conference on Learning Representations (ICLR) , year=

    Mass Editing Memory in a Transformer , author=. The Eleventh International Conference on Learning Representations (ICLR) , year=

  9. [12]

    The Thirteenth International Conference on Learning Representations , year=

    AlphaEdit: Null-Space Constrained Model Editing for Language Models , author=. The Thirteenth International Conference on Learning Representations , year=

  10. [13]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  11. [14]

    2025 , eprint=

    Large Language Diffusion Models , author=. 2025 , eprint=

    2025

  12. [15]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  13. [17]

    Correlation Matrix Memories , year=

    Kohonen, Teuvo , journal=. Correlation Matrix Memories , year=

  14. [19]

    Knowledge Editing in Language Models , author=

    Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models , author=. 2023 , eprint=

    2023

  15. [20]

    2024 , eprint=

    Is Bigger Edit Batch Size Always Better? -- An Empirical Study on Model Editing with Llama-3 , author=. 2024 , eprint=

    2024

  16. [21]

    2023 , eprint=

    Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors , author=. 2023 , eprint=

    2023

  17. [22]

    2024 , eprint=

    PMET: Precise Model Editing in a Transformer , author=. 2024 , eprint=

    2024

  18. [23]

    2026 , eprint=

    Diffusion-Inspired Masked Fine-Tuning for Knowledge Injection in Autoregressive LLMs , author=. 2026 , eprint=

    2026

  19. [24]

    2025 , eprint=

    LongLLaDA: Unlocking Long Context Capabilities in Diffusion LLMs , author=. 2025 , eprint=

    2025

  20. [25]

    2025 , eprint=

    Masks Can Be Distracting: On Context Comprehension in Diffusion Language Models , author=. 2025 , eprint=

    2025

  21. [26]

    2026 , eprint=

    DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders , author=. 2026 , eprint=

    2026

  22. [27]

    International Conference on Learning Representations , year=

    Towards Continual Knowledge Learning of Language Models , author=. International Conference on Learning Representations , year=

  23. [28]

    2020 , eprint=

    Modifying Memories in Transformer Models , author=. 2020 , eprint=

    2020

  24. [29]

    2025 , eprint=

    LLaDA-MoE: A Sparse MoE Diffusion Language Model , author=. 2025 , eprint=

    2025

  25. [31]

    International Conference on Learning Representations , year=

    Fast Model Editing at Scale , author=. International Conference on Learning Representations , year=

  26. [32]

    2022 , eprint=

    Memory-Based Model Editing at Scale , author=. 2022 , eprint=

    2022

  27. [34]

    2026 , url=

    Haewon Park and Sangwoo Kim and Yohan Jo , booktitle=. 2026 , url=

    2026

  28. [35]

    2024 , eprint=

    A Unified Framework for Model Editing , author=. 2024 , eprint=

    2024

  29. [36]

    Anderson

    James A. Anderson. 1972. https://doi.org/10.1016/0025-5564(72)90075-2 A simple neural network generating an interactive memory . Mathematical Biosciences, 14(3):197--220

    work page doi:10.1016/0025-5564(72)90075-2 1972

  30. [37]

    Louis Bethune, David Grangier, Dan Busbridge, Eleonora Gualdoni, Marco Cuturi, and Pierre Ablin. 2025. https://arxiv.org/abs/2502.06042 Scaling laws for forgetting during finetuning with pretraining data injection . Preprint, arXiv:2502.06042

    arXiv 2025

  31. [38]

    Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.522 Editing factual knowledge in language models . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491--6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics

    work page doi:10.18653/v1/2021.emnlp-main.522 2021

  32. [39]

    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 model editing for language models . In The Thirteenth International Conference on Learning Representations

    2025

  33. [40]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 31 others. 2024. https://arxiv.org/abs/2407.21783 The llama 3 ...

    Pith/arXiv arXiv 2024

  34. [41]

    Akshat Gupta, Dev Sajnani, and Gopala Anumanchipalli. 2024. https://arxiv.org/abs/2403.14236 A unified framework for model editing . Preprint, arXiv:2403.14236

    arXiv 2024

  35. [42]

    Thomas Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2023. https://arxiv.org/abs/2211.11031 Aging with grace: Lifelong model editing with discrete key-value adaptors . Preprint, arXiv:2211.11031

    arXiv 2023

  36. [43]

    Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. 2023. https://arxiv.org/abs/2301.04213 Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models . Preprint, arXiv:2301.04213

    arXiv 2023

  37. [44]

    Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun KIM, Stanley Jungkyu Choi, and Minjoon Seo. 2022. https://openreview.net/forum?id=vfsRB5MImo9 Towards continual knowledge learning of language models . In International Conference on Learning Representations

    2022

  38. [45]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. https://doi.org/10.1145/3571730 Survey of hallucination in natural language generation . ACM Comput. Surv., 55(12)

    work page doi:10.1145/3571730 2023

  39. [46]

    Jan-Christoph Kalo and Leandra Fichtel. 2022. Kamel: Knowledge analysis with multitoken entities in language models. In Automated Knowledge Base Construction

    2022

  40. [47]

    Teuvo Kohonen. 1972. https://doi.org/10.1109/TC.1972.5008975 Correlation matrix memories . IEEE Transactions on Computers, C-21(4):353--359

    work page doi:10.1109/tc.1972.5008975 1972

  41. [48]

    Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2024. https://arxiv.org/abs/2308.08742 Pmet: Precise model editing in a transformer . Preprint, arXiv:2308.08742

    arXiv 2024

  42. [49]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. https://doi.org/10.18653/v1/2022.acl-long.229 T ruthful QA : Measuring how models mimic human falsehoods . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214--3252, Dublin, Ireland. Association for Computational Linguistics

    work page doi:10.18653/v1/2022.acl-long.229 2022

  43. [50]

    Xiaoran Liu, Yuerong Song, Zhigeng Liu, Zengfeng Huang, Qipeng Guo, Ziwei He, and Xipeng Qiu. 2025. https://arxiv.org/abs/2506.14429 Longllada: Unlocking long context capabilities in diffusion llms . Preprint, arXiv:2506.14429

    arXiv 2025

  44. [51]

    Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2025. https://arxiv.org/abs/2308.08747 An empirical study of catastrophic forgetting in large language models during continual fine-tuning . Preprint, arXiv:2308.08747

    Pith/arXiv arXiv 2025

  45. [52]

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

    Pith/arXiv arXiv 2023

  46. [53]

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

    2023

  47. [54]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022 a . https://openreview.net/forum?id=0DcZxeWfOPt Fast model editing at scale . In International Conference on Learning Representations

    2022

  48. [55]

    Manning, and Chelsea Finn

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

    arXiv 2022

  49. [56]

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025. https://arxiv.org/abs/2502.09992 Large language diffusion models . Preprint, arXiv:2502.09992

    Pith/arXiv arXiv 2025

  50. [57]

    Xu Pan, Ely Hahami, Jingxuan Fan, Ziqian Xie, and Haim Sompolinsky. 2026. https://arxiv.org/abs/2510.09885 Diffusion-inspired masked fine-tuning for knowledge injection in autoregressive llms . Preprint, arXiv:2510.09885

    Pith/arXiv arXiv 2026

  51. [58]

    Haewon Park, Sangwoo Kim, and Yohan Jo. 2026. https://openreview.net/forum?id=qz3BkyyHWJ SUIT : Knowledge editing with subspace-aware key-value mappings . In The Fourteenth International Conference on Learning Representations

    2026

  52. [59]

    Julianna Piskorz, Cristina Pinneri, Alvaro Correia, Motasem Alfarra, Risheek Garrepalli, and Christos Louizos. 2025. https://arxiv.org/abs/2511.21338 Masks can be distracting: On context comprehension in diffusion language models . Preprint, arXiv:2511.21338

    Pith/arXiv arXiv 2025

  53. [60]

    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, and 25 others. 2025. https://arxiv.org/abs/2412.15115 Qwen2.5 technical report . Preprint, arXiv:2412.15115

    Pith/arXiv arXiv 2025

  54. [61]

    Peng Wang, Ningyu Zhang, Bozhong Tian, Zekun Xi, Yunzhi Yao, Ziwen Xu, Mengru Wang, Shengyu Mao, Xiaohan Wang, Siyuan Cheng, Kangwei Liu, Yuansheng Ni, Guozhou Zheng, and Huajun Chen. 2024 a . https://doi.org/10.18653/v1/2024.acl-demos.9 E asy E dit: An easy-to-use knowledge editing framework for large language models . In Proceedings of the 62nd Annual M...

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

  55. [62]

    Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. 2024 b . https://arxiv.org/abs/2310.16218 Knowledge editing for large language models: A survey . Preprint, arXiv:2310.16218

    arXiv 2024

  56. [63]

    Xu Wang, Bingqing Jiang, Yu Wan, Baosong Yang, Lingpeng Kong, and Difan Zou. 2026. https://arxiv.org/abs/2602.05859 Dlm-scope: Mechanistic interpretability of diffusion language models via sparse autoencoders . Preprint, arXiv:2602.05859

    arXiv 2026

  57. [64]

    Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.632 Editing large language models: Problems, methods, and opportunities . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10222--10240, Singapore. Assoc...

    work page doi:10.18653/v1/2023.emnlp-main.632 2023

  58. [65]

    Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. 2025. Dream 7b: Diffusion large language models. arXiv preprint arXiv:2508.15487

    Pith/arXiv arXiv 2025

  59. [66]

    Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. https://arxiv.org/abs/2012.00363 Modifying memories in transformer models . Preprint, arXiv:2012.00363

    arXiv 2020

  60. [67]

    Fengqi Zhu, Zebin You, Yipeng Xing, Zenan Huang, Lin Liu, Yihong Zhuang, Guoshan Lu, Kangyu Wang, Xudong Wang, Lanning Wei, Hongrui Guo, Jiaqi Hu, Wentao Ye, Tieyuan Chen, Chenchen Li, Chengfu Tang, Haibo Feng, Jun Hu, Jun Zhou, and 7 others. 2025. https://arxiv.org/abs/2509.24389 Llada-moe: A sparse moe diffusion language model . Preprint, arXiv:2509.24389

    arXiv 2025