Pith. sign in

REVIEW 2 major objections 7 minor 1 cited by

Is Fine-Tuning an Effective Solution? Reassessing Knowledge Editing for Unstructured Data

T0 review · 2 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning with the right loss scope, layer, component, and chat template outperforms dedicated locate-and-edit methods for unstructured knowledge editing, and its advantage grows with batch size from +6.78% to +10.80%.

desk verdict Worth engaging: solid new locality benchmarks and a fine-tuning recipe that beats locate-and-edit, but the batch-trend claim is cherry-picked across baselines. read the letter →

arxiv 2506.09672 v1 pith:GAVNL6QR submitted 2025-06-11 cs.CL cs.AI

Hao Xiong , Chuanyuan Tan , Wenliang Chen This is my paper
classification cs.CLcs.AI
keywords knowledgeeditingunstructuredfine-tuninglocalityevaluationbatchlargelanguagemodelslow-rankadaptationlocate-and-edit
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the reported failure of fine-tuning (FT) in unstructured knowledge editing (UKE) is an artifact of suboptimal configurations rather than an inherent limitation. It constructs two locality-augmented UKE datasets, UnKEBench-Loc and AKEW-Loc (CF), and identifies four factors—loss calculation scope, layer selection, component selection, and chat template—that determine how well FT-based methods edit long, free-form knowledge. With the best settings (FT-UKE: loss over all target tokens, single middle layer, down-projection weights, chat template), fine-tuning consistently beats the previous state-of-the-art UKE methods on both datasets and both tested models. In batch editing, FT-UKE's average overall-score lead over SOTA grows from +6.78% to +10.80% as batch size increases from 1 to 100. If correct, this shifts the default baseline for UKE research and pushes locality evaluation beyond a single multiple-choice set like MMLU.

What carries the argument

The carrying mechanism is a four-factor recipe for fine-tuning in UKE. (1) Loss calculation scope: computing the loss over all target tokens rather than only the final prediction token; (2) layer selection: editing a single middle transformer layer for direct weight fine-tuning, versus all layers for adapter-based tuning; (3) component selection: editing the FFN down-projection (downproj) for FT-UKE and all attention projections for AdaLoRA-UKE; and (4) chat template: applying the model's chat template during editing. The two constructed datasets, UnKEBench-Loc and AKEW-Loc (CF), supply locality test data of three types—related Wikipedia documents, random Wikipedia documents, and structured triples from KnowEdit—so that post-edit knowledge preservation is measured directly instead of through a small multiple-choice set.

What would settle it

A held-out validation test: select FT-UKE's four factor settings using only UnKEBench-Loc, then evaluate on AKEW-Loc (CF) (or on a third UKE dataset such as MQUAKE-CF or WikiUpdate), and check whether the overall-score lead over UnKE and AnyEdit still holds. If the advantage shrinks to near zero or reverses, the claim of FT-based superiority over SOTA is an artifact of tuning on the test set.

Watch

Extended reading notes

Core claim

The central discovery is that a fine-tuned baseline, configured deliberately, is stronger than specialized locate-and-edit methods for unstructured knowledge editing. The paper shows that the best fine-tuning variant, FT-UKE, achieves higher overall scores (average of Edit Success, Generalization, and Locality) than UnKE and AnyEdit, the dedicated UKE methods, on UnKEBench-Loc and AKEW-Loc (CF) across Llama3-8B-Instruct and Qwen2.5-7B-Instruct, with one exception in the Bert Score of overall score on AKEW-Loc (CF) for Llama3. The performance gap widens under batch editing, where FT-UKE degrades more slowly than its competitors as batch size grows. The paper further demonstrates that locality evaluation on MMLU multiple-choice questions compresses all methods within about ±1.6% of the pre-edit model, whereas the newly built locality data separates methods by much wider margins, supporting the need for multi-type locality tests.

Load-bearing premise

The best factor settings were chosen by maximizing overall score on AKEW-Loc (CF) in Section 5.3 and then evaluated on that same dataset in Section 5.2, so the reported SOTA advantage assumes this selection does not inflate results and that the recipe generalizes to other datasets and models.

Editorial extensions

If this is right

  • Subsequent UKE papers should adopt FT-UKE as the default baseline instead of the previously reported FT variants; comparisons against it are needed to claim improvement.
  • Batch editing should be added to standard UKE evaluation, since method rankings at batch size 1 do not match rankings at larger sizes (e.g., AnyEdit falls behind UnKE at batch 50).
  • Locality evaluation for UKE should use the multi-type test data (related document, random document, structured triple) to distinguish methods that MMLU-Loc cannot resolve.
  • The recipe transfers across at least two instruction-tuned 7B–8B models, suggesting it may hold for similarly styled chat models, though this is tested only on Llama3-8B-Instruct and Qwen2.5-7B-Instruct.

Reading between the lines

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

  • If FT-UKE's advantage is robust across more models and domains, the field's locate-and-edit paradigm may be over-engineered: general optimization with the right inductive biases could be the simpler explanation for earlier successes attributed to causal intervention.
  • The four-factor recipe is likely to interact with model family, scale, and chat-template style; a natural extension would be a systematic sweep over more base models and template variants to map where the advantage persists.
  • The locality datasets could be extended back to structured editing: using the same RelDoc/RandDoc/StructTrip construction with SKE queries would give a common locality benchmark across both editing paradigms.
  • A testable prediction from the batch-editing result: FT-UKE's gradual degradation suggests weight-space updates are less batch-sensitive than activation-based edits; comparing with methods that constrain activations at larger batch sizes could reveal which component drives the effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. This paper reassesses fine-tuning (FT) based methods for unstructured knowledge editing (UKE). The authors construct two locality-augmented datasets, UnKEBench-Loc and AKEW-Loc (CF), by extending existing UKE datasets with three types of locality data (related documents, random documents, and structured triples). They identify four factors affecting FT-based editing (loss calculation scope, layer selection, component selection, and chat template) and, through factor analysis on AKEW-Loc (CF), derive a recipe (FT-UKE) that they claim outperforms state-of-the-art UKE methods in single and batch editing scenarios, with the batch advantage growing from +6.78% to +10.80% as batch size increases. They also argue that MMLU-based locality evaluation is insufficient and that their datasets provide more sensitive locality assessment. The paper includes detailed results on two models (Llama3-8B-Instruct and Qwen2.5-7B-Instruct) and promises code and data release.

Significance. If the empirical claims are sound, the paper makes a valuable contribution: it challenges the prevailing assumption that fine-tuning is ineffective for UKE, provides a simple and reproducible training recipe, and introduces locality benchmarks that are more discriminative than existing practice. The factor analysis is a useful reference for future work, and the construction of the datasets is a contribution in its own right. However, the two headline empirical claims—the SOTA advantage on AKEW-Loc (CF) and the monotonic increase of the batch-editing advantage—rest on methodological choices that need to be addressed. The paper is clearly written and the tables are detailed, but the central numbers as currently presented are not fully trustworthy.

major comments (2)
  1. [5.2/5.3, Tables 3 and 4] The FT-UKE configuration reported as state-of-the-art in Section 5.2 was selected on the same benchmark that is used to establish the headline result. In Section 5.3, Table 4, the authors evaluate factor combinations by editing Llama3-8B-Instruct on AKEW-Loc (CF) and choose the row 'single downproj' with 'all target tokens' as FT-UKE (highlighted in green). This exact configuration then appears in Table 3 as the SOTA result on AKEW-Loc (CF), with the reported +6.78% average advantage over AnyEdit at batch size 1. No validation split, cross-validation, or held-out portion of AKEW-Loc (CF) is described. Since the configuration was chosen to maximize OA on this dataset, the advantage on AKEW-Loc (CF) is inflated by selection on the test set. The configuration also wins on UnKEBench-Loc (Table 3), which provides some independent support, but the paper should either (a) select the recipe on one dataset and test on the other, (b) use a held-out split of AKEW-Loc (CF) for model selection, or (c) explicitly report the selection procedure and re-estimate the advantage without selection. As written, the claim that FT-UKE 'consistently outperforms the SOTA UKE methods' is overstated for the tuned benchmark.
  2. [5.4, Table 7, Figure 3, Abstract] The batch-editing advantage claim is not supported by the reported numbers. The abstract and Section 5.4 state that FT-UKE's average OA lead over SOTA UKE methods grows from +6.78% (batch size 1) to +10.80% (batch size 100). These two endpoints are computed against different baselines: +6.78% is the average lead over AnyEdit at BS=1 (Table 7: 85.09-80.65 for BS and 71.34-62.22 for RL, mean 6.78), whereas +10.80% is the average lead over UnKE at BS=100 (81.54-74.37 for BS and 61.60-47.17 for RL, mean 10.80). Against a fixed baseline the trend is not monotonic: versus UnKE the lead is 9.36, 9.36, 6.52, 10.80 at BS=1,10,50,100, and versus AnyEdit it is 6.78, 7.05, 18.16, 18.85. Even when using the best prior UKE method at each batch size (AnyEdit at BS=1 and 10, UnKE at BS=50 and 100), the lead drops from 7.05 at BS=10 to 6.52 at BS=50 before rising to 10.80 at BS=100. The claim that FT-UKE's advantage 'increases as the batch size grows' is therefore not supported by the paper's own tables. The authors should recompute the advantage against a fixed baseline (or the best baseline at each size) and either revise the claim or report a statistically sound measure of the trend.
minor comments (7)
  1. [3, Appendix A] The retrieval cutoff K for RelDoc is never specified. The text and Figure 2 refer to 'Top K' but no value is given; please report K and the number of retrieved documents per query, as this affects dataset reproducibility and the 'factually disjoint' filtering.
  2. [5.3, Table 4 and Appendix C, Table 6] The main-text discussion refers to 'the best settings ... are highlighted in green' and says all settings in Table 4 apply the chat template, but this is only explicit in the Appendix. Add a note under Table 4 that all rows use the chat template so readers do not misread the factor comparison.
  3. [5.4, Figure 3] The figure legend and text refer to 'SOTA UKE methods,' but the figure does not identify which baseline is SOTA at each batch size (AnyEdit at BS=1 and 10, UnKE at BS=50 and 100). Please mark the SOTA baseline in the figure or in the caption.
  4. [5.5, Table 5] The comparison between MMLU-Loc (accuracy) and the proposed locality data (BERTScore/ROUGE-L) mixes different metric types; the 'minimum gap' argument should be framed as a qualitative difference rather than a direct measure of sensitivity, and the table should label the metric columns accordingly.
  5. [5.2, Table 3] Several cells are footnoted with + and * and the table is difficult to parse; please separate cited baselines from runs performed in this work, and consider reporting standard deviations or significance tests for the key differences (e.g., FT-UKE vs AdaLoRA-UKE on AKEW-Loc (CF) BS OA for Llama3), since some margins are small.
  6. [Limitations] The statement that further exploration is unnecessary because 'FT-UKE in the settings we experimented with already surpasses the existing SOTA methods' relies on the same test-set selection issue raised in Major Comment 1; please qualify this claim or reference the validation strategy.
  7. [Throughout] Typographical and consistency issues: 'Batch Ediging' in the Appendix D heading should be 'Batch Editing'; 'such us UnKE' in Section 2.1 should be 'such as UnKE'; 'UNKEBENCH-LOC' in the Appendix E table should be 'UnKEBench-Loc'; and 'Detailed results are listed in Appendix (Table 7)' should be 'Appendix D (Table 7)'.

Circularity Check

1 steps flagged · score 6.0 of 10

FT-UKE is selected by maximizing OA on AKEW-Loc (CF) and then reported as SOTA-beating on that same dataset, a partial selection-on-evaluation-set circularity; UnKEBench-Loc and Qwen results provide independent evidence.

  1. fitted input called prediction [Section 5.3 / Table 4; Section 5.2 / Table 3]
    "In this section, we edit Llama3-8B-Instruct on AKEW-Loc (CF) using FT-based methods, applying the factor settings discussed in § 4. ... The best results for each group are highlighted in bold, and the settings used in §5.2 are highlighted in green (FT-UKE, AdaLoRA-UKE). ... Compared to the best UKE method in the previous studies, AnyEdit, it exceeds by 4.44% and 9.12% in the BS and RL of OA score on AKEW-Loc (CF), with an average advantage of 6.78%."

    The FT-UKE configuration is not a fixed baseline; it is the green-highlighted row of Table 4 that achieves the highest OA on AKEW-Loc (CF) among the tested factor combinations. Section 5.2 then reports exactly that row's OA (85.09/71.34) as the headline result on AKEW-Loc (CF) and derives the +6.78% advantage over AnyEdit from it. Thus the evaluation set was used to select the recipe (loss scope, layer, component, chat template), and the same evaluation set is then used to confirm that the recipe beats SOTA. The AKEW-Loc (CF) superiority claim is therefore a selection artifact rather than an independent test.

full rationale

The load-bearing circular step is the AKEW-Loc (CF) main result. Section 5.3 states that the factor analysis is conducted by editing Llama3-8B-Instruct on AKEW-Loc (CF), and Table 4 marks the settings used in Section 5.2 in green. Those green rows are the best-OA rows in their groups, so FT-UKE and AdaLoRA-UKE are defined as the best-performing configurations on AKEW-Loc (CF). Section 5.2 then uses those same configurations to claim that FT-UKE consistently outperforms SOTA UKE methods on AKEW-Loc (CF) and to compute the +6.78% average advantage. Numerically, the reported FT-UKE OA on AKEW-Loc (CF) is identical to the selected green row in Table 4, so the AKEW-Loc (CF) comparison is selection-on-the-evaluation-set: the advantage is partly built into the factor choice rather than being an out-of-sample prediction. The paper is not wholly circular because FT-UKE with the selected recipe is also evaluated on UnKEBench-Loc and on Qwen2.5-7B-Instruct, where the recipe was not tuned, and those results independently support the general conclusion. The batch-editing claim that the advantage grows from +6.78% to +10.80% compares different baselines at different batch sizes, which is an internal-consistency issue rather than a circularity; it is therefore not scored as a separate circular step. Overall, one central reported prediction reduces by construction on AKEW-Loc (CF), while substantial independent content remains.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests less on mathematical derivation than on the validity of the new benchmarks and the choice of the winning fine-tuning configuration. The largest unstated burden is the assumption that selecting settings on AKEW-Loc (CF) and then reporting results on the same set does not inflate the headline comparison.

free parameters (2)
  • FT-UKE factor configuration = loss on all target tokens; single layer; downproj; with chat template
    Selected by comparing factor variants on AKEW-Loc (CF) in Section 5.3, then used in the main comparisons in Section 5.2 on the same dataset; this is a post hoc choice rather than a fixed a priori rule.
  • RelDoc top-K retrieval cutoff = not reported
    Figure 2 shows 'Top K' for relevant document retrieval, but the text never gives K; this affects which documents are labeled semantically related and therefore the locality scores.
assumptions (3)
  • domain assumption RelDoc documents, after OpenIE filtering, are semantically related to the edit query but factually disjoint from it.
    The locality claim for RelDoc depends on this disjointness; if the filtered documents still contain the edited fact, a drop in score may reflect edit failure rather than locality damage. See Section 3 and Appendix A.
  • domain assumption BERTScore and ROUGE-L computed with all-MiniLM-L6-v2 measure knowledge preservation well enough to compare methods.
    All metrics in Tables 3, 4, and 5 rely on these similarities; a different metric could reorder methods. Mentioned in Section 5.1 footnote 3.
  • ad hoc to paper The factor configuration selected on AKEW-Loc (CF) transfers to other datasets and models.
    Section 5.3 picks the best settings on AKEW-Loc (CF) without a held-out validation split, and Section 5.2 then applies them to UnKEBench-Loc and to Qwen2.5; the transfer is assumed, not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is Fine-Tuning an Effective Solution? Reassessing Knowledge Editing for Unstructured Data." pith.science (2026). https://pith.science/paper/GAVNL6QR

@misc{pith2026250609672,
  author       = {Pith},
  title        = {Pith review of: Is Fine-Tuning an Effective Solution? Reassessing Knowledge Editing for Unstructured Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GAVNL6QR}},
  note         = {Machine review of arXiv:2506.09672}
}
read the original abstract

Unstructured Knowledge Editing (UKE) is crucial for updating the relevant knowledge of large language models (LLMs). It focuses on unstructured inputs, such as long or free-form texts, which are common forms of real-world knowledge. Although previous studies have proposed effective methods and tested them, some issues exist: (1) Lack of Locality evaluation for UKE, and (2) Abnormal failure of fine-tuning (FT) based methods for UKE. To address these issues, we first construct two datasets, UnKEBench-Loc and AKEW-Loc (CF), by extending two existing UKE datasets with locality test data from the unstructured and structured views. This enables a systematic evaluation of the Locality of post-edited models. Furthermore, we identify four factors that may affect the performance of FT-based methods. Based on these factors, we conduct experiments to determine how the well-performing FT-based methods should be trained for the UKE task, providing a training recipe for future research. Our experimental results indicate that the FT-based method with the optimal setting (FT-UKE) is surprisingly strong, outperforming the existing state-of-the-art (SOTA). In batch editing scenarios, FT-UKE shows strong performance as well, with its advantage over SOTA methods increasing as the batch size grows, expanding the average metric lead from +6.78% to +10.80%

Figures

Figures reproduced from arXiv: 2506.09672 by the authors.

Figure 1
Figure 1. Comparison between structured and unstruc [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Data Collection Process. We sample unstruc [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. OA of different methods for editing Llama3-8B-Instruct on AKEW-Loc (CF) in batch editing scenarios. FT [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Implicit Reasoning Steering via Concept Chaining

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Reinforcement-learning-optimized concept-chain paragraphs covertly steer language-model multiple-choice preferences after continued pretraining, with far lower detectability than direct paraphrases.

Reference graph

Works this paper leans on

31 extracted references · 6 canonical work pages · cited by 1 Pith paper

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

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

    2024

  3. [3]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, and 1 others. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609

    arXiv 2023

  4. [4]

    Steven Bird and Edward Loper. 2004. https://aclanthology.org/P04-3031/ NLTK : The natural language toolkit . In Proceedings of the ACL Interactive Poster and Demonstration Sessions , pages 214--217, Barcelona, Spain. Association for Computational Linguistics

    work page 2004

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  6. [6]

    Jingcheng Deng, Zihao Wei, Liang Pang, Hanxing Ding, Huawei Shen, and Xueqi Cheng. 2024. Everything is editable: Extend knowledge editing to unstructured data in large language models. arXiv preprint arXiv:2405.15349

    arXiv 2024

  7. [7]

    Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, and Tat-Seng Chua. 2024. Alphaedit: Null-space constrained knowledge editing for language models. arXiv preprint arXiv:2410.02355

    arXiv 2024

  8. [8]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. https://openreview.net/forum?id=d7KBjmI3GmQ Measuring massive multitask language understanding . In International Conference on Learning Representations

    2021

Show all 31 references
  1. [9]

    Chenhui Hu, Pengfei Cao, Yubo Chen, Kang Liu, and Jun Zhao. 2024. https://doi.org/10.18653/v1/2024.findings-acl.207 W il KE : Wise-layer knowledge editor for lifelong knowledge editing . In Findings of the Association for Computational Linguistics: ACL 2024, pages 3476--3503, ...

    2024 doi

  2. [10]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

    2022

  3. [11]

    Houcheng Jiang, Junfeng Fang, Ningyu Zhang, Guojun Ma, Mingyang Wan, Xiang Wang, Xiangnan He, and Tat-seng Chua. 2025. Anyedit: Edit any knowledge encoded in language models. arXiv preprint arXiv:2502.05628

    2025

  4. [12]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.550 Dense passage retrieval for open-domain question answering . In Proceedings of the 2020 Conference on Empiric...

    2020 doi

  5. [13]

    Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2024. Pmet: Precise model editing in a transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18564--18572

    2024

  6. [14]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013/ ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  7. [15]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in neural information processing systems, 35:17359--17372

    2022

  8. [16]

    Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. 2023. Mass-editing memory in a transformer. In ICLR

    2023

  9. [17]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022. Fast model editing at scale. In International Conference on Learning Representations

    2022

  10. [18]

    Team Qwen. 2024. https://qwenlm.github.io/blog/qwen2.5/ Qwen2.5: A party of foundation models

    2024

  11. [19]

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

    2024

  12. [20]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  13. [21]

    Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. 2024 a . Wise: Rethinking the knowledge memory for lifelong model editing of large language models. Advances in Neural Information Processing Systems, 37:53764--53797

    2024

  14. [22]

    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 b . https://doi.org/10.18653/v1/2024.acl-demos.9 E asy E dit: An easy-to-use knowledge e...

    2024 doi

  15. [23]

    Xiaobao Wu, Liangming Pan, William Yang Wang, and Anh Tuan Luu. 2024. Akew: Assessing knowledge editing in the wild. arXiv preprint arXiv:2402.18909

    2024 arXiv

  16. [24]

    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 Empiri...

    2023 doi

  17. [25]

    Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, and 1 others. 2024. A comprehensive study of knowledge editing for large language models. arXiv preprint arXiv:2401.01286

    2024 arXiv

  18. [26]

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512

    2023 arXiv

  19. [27]

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

    2019 arXiv

  20. [28]

    Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.296 Can we edit factual knowledge by in-context learning? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Pro...

    2023 doi

  21. [29]

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

    2020 arXiv

  22. [30]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  23. [31]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.