Pith. sign in

REVIEW 3 major objections 6 minor 20 references

ChainEdit: Propagating Ripple Effects in LLM Knowledge Editing through Logical Rule-Guided Chains

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ChainEdit turns a single knowledge edit into a logically linked batch edit: on the RIPPLE EDITS benchmark, logical generalization for MEMIT on Llama-3-8B rises from 18.6% to 58.7%, with comparable gains for other edit methods and a…

desk verdict The 40-point LG jump is likely answer-injection: ChainEdit edits the benchmark's own expected answers, so the paper's central claim about logical generalization is untested. read the letter →

arxiv 2507.08427 v1 pith:L7DZ2WW6 submitted 2025-07-11 cs.CL cs.AI

Zilu Dong , Xiangqing Shen , Zinong Yang , Rui Xia This is my paper
classification cs.CLcs.AI
keywords knowledgeeditinglogicalgeneralizationrippleeffectsrulemininggraphsLLMreasoningbatchevaluationbias
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

The paper is trying to establish that LLM knowledge editing fails on ripple effects because it edits one fact in isolation, leaving logically tied facts frozen, and that this failure can be fixed by editing a whole chain of related facts at once. The proposed framework, ChainEdit, mines candidate logical rules from a knowledge graph (nationality follows from birthplace plus city-of, for example), asks the LLM which rules it judges universal, converts the survivors into directive rules, and at edit time uses them to generate derived facts that are edited together with the original fact. On the ripple-effect benchmark, this lifts logical generalization for MEMIT on Llama-3-8B from 18.6% to 58.7%, with similar gains for other edit methods and for a 1.5B-parameter model, while the original edit still lands and unrelated knowledge is not broadly disturbed. The paper also argues that existing benchmarks conflate logical generalization with the LLM agreeing with an external knowledge graph, and it proposes filtered, replaced, and in-prompt variants that separate the two. If the framework is right, rule-guided chain editing is a model-agnostic way to make knowledge edits logically coherent without retraining.

What carries the argument

The load-bearing object is the directive rule, written $\langle \phi, \psi \rangle$: $\phi$ is the trigger relation of the edit, and $\psi=(s_{\text{new}}, r_{\text{new}}, o_{\text{new}})$ is a knowledge-generation template with entity placeholders $S$ and $O$ and dot-operator multi-hop queries such as $O.\text{spouse}$ or $S.\text{birthplace}.\text{country}$. A mined rule such as "mother follows from father and spouse" is expanded into one or more directive rules, and ambiguous rules explicitly encode every legitimate update path. At edit time, the trigger relation selects the rules, variable substitution grounds the templates in the edited entities, the LLM answers prompted questions such as "the spouse of Carol is" to fill in missing entities, and the resulting triples are added to the edit batch. This machinery converts one weight-level edit into a logically closed cluster of edits, which is why the paper calls the framework a chain edit.

What would settle it

Build an evaluation set whose answer chains are generated from a knowledge graph disjoint from the one used to mine ChainEdit's rules, or from human-written logical constraints, and re-run the same edited facts; if logical generalization then falls back toward the roughly 20% baseline while edit reliability stays high, the reported gain is mostly an artifact of rule/benchmark overlap rather than general logical propagation.

Watch

Extended reading notes

Core claim

The paper's central claim is that the missing ingredient in knowledge editing is an explicit, machine-readable link between the edited fact and the facts that must change with it. ChainEdit supplies that link in four phases: mine multi-hop relational paths from a knowledge graph as candidate rules; filter the rules by asking the LLM to judge their universality, keeping only those the model itself endorses; convert each surviving rule into a directive rule with a trigger relation and a template for generating a derived triple, including templates that encode both valid update paths when a rule is ambiguous; and at edit time batch-edit the original triple together with the derived triples. The headline result is that MEMIT's logical generalization on the RIPPLE EDITS benchmark rises from 18.6% to 58.7% on Llama-3-8B, and the same pattern appears for other edit methods and for a smaller model. The paper concludes that the LLM's own reasoning is essential to the pipeline: adding the LLM-alignment filter improves logical generalization by roughly seven points over frequency-based mining alone, whereas human curation of the final rule set adds little.

Load-bearing premise

The result rests on assuming that high-frequency paths in the knowledge graph used for mining are the right logical rules to test, and that the benchmark built from the same family of knowledge graphs is an unbiased test of those rules; if an independent source of logic were used to evaluate the edits, the large improvement could shrink.

Editorial extensions

If this is right

  • Wrapping an existing edit method with ChainEdit roughly triples logical generalization without changing the method's own update rule, so the gain is available to any batch-edit method that can accept multiple triples.
  • The original edit's success rate is preserved: when MEMIT is replaced by the variant that avoids same-subject conflicts, reliability stays near 98%, and the disturbance to unrelated knowledge moves only by a few percentage points.
  • The gain survives when intermediate knowledge is filtered out, replaced by the LLM's own knowledge, or placed directly in the prompt, so the improvement is not merely alignment with the external knowledge graph used to build the benchmark.
  • Applying the rules to a batch of edits continues to help at sizes up to 100, though the logical-generalization advantage narrows slightly as edit batches grow and the underlying edit method itself degrades.

Reading between the lines

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

  • Editorial extension: the rule set and the benchmark are both derived from knowledge graphs, so the cleanest stress test is to build a benchmark from a disjoint knowledge source; on that test the 40-point gain could shrink if part of it reflects rule/benchmark overlap rather than genuine logical propagation.
  • Editorial extension: because derived triples are filled in by the LLM before editing, a wrong intermediate answer (for example, the wrong spouse) becomes an edited falsehood; the framework invites a confidence threshold or multi-sample voting before committing derived triples to the edit batch.
  • Editorial extension: the LLM-alignment filter inherits the judge model's own blind spots, so weaker models may endorse overbroad rules; one could test this by using a stronger judge model to select rules and measuring whether the smaller model's logical generalization score improves further.
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

3 major / 6 minor

Summary. The paper proposes ChainEdit, a plug-in framework for LLM knowledge editing that (1) mines logical rules from Wikidata for relations appearing in the RIPPLE EDITS POPULAR benchmark, (2) filters and aligns those rules with LLM judgments, (3) for each edit (s, r, o) generates derived triples such as (s, r_new, LLM_answer) by querying the LLM along rule paths, and (4) batch-edits the original and derived triples with MEMIT, MEMIT-Merge, FT-M, or LoRA. On RIPPLE EDITS, the authors report large gains in logical generalization (e.g., MEMIT on Llama-3-8B: from 18.6% to 58.7%) and introduce three dataset variants (filtered, replaced, in-prompt) intended to reduce dependence on external knowledge-graph facts during evaluation.

Significance. If the central claim were established, ChainEdit would be a practical, model-agnostic way to propagate knowledge edits along relational chains, and the proposed dataset variants would be a useful diagnostic toolkit for separating logical reasoning from external-knowledge dependence. The paper's strengths include experimental breadth (two models, four editing methods, single- and batch-instance settings), an explicit attempt to align symbolic rules with LLM reasoning, and a code-release plan. The weakness is that the headline LG metric appears to measure whether the pipeline wrote the benchmark's expected answers into the weights, not whether the edited model can infer related facts; the proposed dataset variants do not remove this confound.

major comments (3)
  1. [3.3; 4.1.1; Table 1] The headline LG gains do not establish logical generalization because the evaluation is circular. Section 3.1.1 mines rules from Wikidata; Section 4.1.1 evaluates on RIPPLE EDITS POPULAR, whose LG questions are constructed from the same kind of KG-derived relational dependencies; Section 3.3 then queries the LLM for the object of the related relation (e.g., 'spouse of Carol') and batch-edits the resulting triple into the model. The LG metric asks exactly for such related facts, so a correct answer can be produced by the edit operation itself rather than by logical generalization from the edited fact. The paper needs at least one of the following controls: (a) rules mined from an independent KG (e.g., YAGO or ConceptNet) evaluated on POPULAR; (b) a held-out set of relational paths or entities not seen during rule mining; or (c) a sanity baseline that skips rule mining and directly asks the LLM for the related fact and then edits it. Without such a control, the 40-point improvement cannot be attributed to logical generalization.
  2. [4.4; Tables 3, 5, 6, 7] The filtered/replaced/in-prompt dataset variants address a real concern about external-knowledge dependence, but they do not resolve the answer-injection issue, because ChainEdit continues to edit the specific triples that the LG questions later probe. In the filtered variant, the inclusion criterion (questions answerable with the model's internal knowledge) may even make the LLM query in Section 3.3 more likely to return exactly the benchmark's expected answer; the 71.0% LG in Table 3 is therefore not a cleaner measure of generalization. Please report, for each variant, what happens when the derived triple is excluded from the edit batch; this single ablation would separate 'the model inferred the related fact' from 'the answer was explicitly written in.'
  3. [Abstract; Table 1; 4.3] The abstract's claim of 'preserving editing reliability' is contradicted by Table 1: MEMIT reliability drops from 99.8% to 90.0% on Llama-3-8B and from 99.0% to 87.4% on Qwen2.5-1.5B, and Section 4.3 explicitly acknowledges the 9.8-point drop on Llama. Forgetfulness also increases in several configurations (e.g., MEMIT on Llama: 31.2 to 37.0; LoRA on Qwen: 19.4 to 24.1). The method may still be useful, but the reliability claims should be softened and supported by a comparison against a same-batch baseline that edits an equal number of random or unrelated triples.
minor comments (6)
  1. [4.3.1; Table 2] The text says the 7% improvement (53% to 60%) and the 60.0% versus 61.5% comparison are for Llama-3-8B with MEMIT-Merge, but Table 2 reports only Qwen2.5-1.5B and the cited numbers match the Qwen MEMIT-Merge rows; please correct the model labels.
  2. [3.3] The notation 'Ri whose ϕi = r' is ambiguous because ϕ is introduced as a logical predicate while r is a relation name; please use distinct names for the relation and the predicate placeholder throughout.
  3. [Tables 1-7] The tables report single-run percentages without standard errors, confidence intervals, or dataset sizes; given the large reported differences this is not fatal, but reporting variance or at least the number of edits would strengthen the claims.
  4. [Equation (1)] Equation (1) uses Alice and Carol without connecting them to the preceding example's entities A and B; please align the entity names for readability.
  5. [4.3.1] The rule-set calibration step ('calibrate both rule sets to comparable sizes, around 180 rules') is not specified; please report the threshold γ, the number of rules before and after each stage, and the number of human-curated rules.
  6. [4.4] The phrase 'maintains question quantity ( vs. reduction in filtered data)' appears to contain a typo; please clarify the intended comparison.

Circularity Check

1 steps flagged · score 8.0 of 10

LG gains are produced by batch-editing the benchmark answers themselves, so the logical-generalization claim reduces by construction.

  1. fitted input called prediction [Section 3.3 (Rule Application) and Section 4.1.1 (Existing Datasets)]
    "Through variable substitution, we obtain the query item q = (Alice, mother, Carol.spouse), then query LLM to get the entity representing Carol.spouse. ... This prompt is then fed to the LLM to elicit a response vj, and to finally formulate a new knowledge triple (sj, rj, vj) ... Eventually, we perform batch editing on both original and derived knowledge using existing knowledge editing methods. ... Among these, LG, CI, CII, and SA assess the model’s ability to propagate edits to logically related knowledge."

    The LG metric asks whether the model outputs the logically related fact after an edit. ChainEdit constructs exactly that fact: it queries the LLM for the object of the related relation (e.g., 'spouse of Carol' -> Mary), forms the triple (Alice, mother, Mary), and batch-edits it together with the original edit. A correct LG answer is therefore not evidence that the model generalized the edited fact through a learned rule; it is evidence that the separately edited triple was written into the weights and retrieved. The reported gain (18.6% to 58.7% for MEMIT) measures the success of the extra direct edit, not logical generalization. No control that edits the LLM's answer without rule mining is provided, so the central claim is untested.

full rationale

The headline LG improvement is circular by construction: the 'derived knowledge' that ChainEdit batch-edits is the same logically related fact that the LG metric checks. Section 3.3 generates (sj, rj, vj) by prompting the LLM, then edits it; Section 4.1.1 defines LG as propagating edits to logically related knowledge. Thus LG after ChainEdit is essentially the reliability/accuracy of an extra direct edit, not a measure of logical generalization. The paper's dataset variants (filtered, replaced, in-prompt) address external-knowledge mismatches but do not remove this answer-injection: in all variants, the final answer is still written by the batch edit rather than inferred from the edited fact. Consequently, the central claim of a 30%+ logical generalization gain reduces to the fact that the method edits the benchmark's own target answers. Other reported metrics (reliability, specificity, forgetfulness) are not circular and retain independent value, but the paper's primary contribution is the LG gain, which is forced by the method-evaluation alignment. Hence score 8, not 10, because the rule-mining and alignment components are real and the non-LG metrics are independently measured.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The framework rests on five free choices and five unproven assumptions. The most consequential is the domain assumption that Wikidata frequency equals logical validity, combined with the benchmark also being KG-derived, which creates the circularity burden described in the circularity score.

free parameters (5)
  • Support threshold gamma for rule mining = not reported
    Controls which inverse relations and multihop paths are retained as candidate rules in Section 3.1.1; the paper does not state the numerical value or how it was chosen.
  • Rule set size after calibration = about 180 rules
    Section 4.3.1 calibrates the pure-KG and LLM-aligned rule sets to comparable sizes, but 180 is a manually chosen size, not derived from any principle.
  • Per-relation sample count = 10,000
    Used for inverse and multihop path mining in Section 4.2.1; affects which rules are discovered.
  • LLM alignment acceptance thresholds = True and Usually True accepted
    Appendix B accepts only the top two of five confidence levels as valid rules; the cutoff is a free choice.
  • Human curation effort = not quantified
    Section 4.3.1 mentions manual curation in the full pipeline for computational efficiency, but the number of human-reviewed rules and selection criteria are not specified.
assumptions (5)
  • domain assumption High-frequency relational paths in Wikidata are valid logical rules for LLM knowledge editing.
    Invoked in Section 3.1.1; frequency in the KG is treated as logical validity, without a separate validity check beyond LLM self-report.
  • domain assumption LLM self-reported confidence reliably identifies universally valid rules.
    Section 3.1.2 and Appendix B use the LLM's True and Usually True ratings to filter rules; there is no external verification of these judgments.
  • domain assumption The RIPPLE EDITS POPULAR benchmark ground truth is the correct target for logical generalization.
    Section 4.1.1 adopts this benchmark without questioning whether its KG-derived answers are the only acceptable logically consistent answers.
  • ad hoc to paper The filtered, replaced, and in-prompt variants measure logical generalization rather than selecting easy cases.
    Section 4.1.2 removes questions the LLM cannot answer from internal knowledge, which may preferentially remove failures of the proposed method.
  • domain assumption Batch-editing the original and derived triples with the same edit method preserves the reliability of the original edit.
    Section 3.3 assumes the base editor can safely handle multiple edits; the paper's own Table 1 shows this fails for MEMIT on Llama.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ChainEdit: Propagating Ripple Effects in LLM Knowledge Editing through Logical Rule-Guided Chains." pith.science (2026). https://pith.science/paper/L7DZ2WW6

@misc{pith2026250708427,
  author       = {Pith},
  title        = {Pith review of: ChainEdit: Propagating Ripple Effects in LLM Knowledge Editing through Logical Rule-Guided Chains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7DZ2WW6}},
  note         = {Machine review of arXiv:2507.08427}
}
read the original abstract

Current knowledge editing methods for large language models (LLMs) struggle to maintain logical consistency when propagating ripple effects to associated facts. We propose ChainEdit, a framework that synergizes knowledge graph-derived logical rules with LLM logical reasoning capabilities to enable systematic chain updates. By automatically extracting logical patterns from structured knowledge bases and aligning them with LLMs' internal logics, ChainEdit dynamically generates and edits logically connected knowledge clusters. Experiments demonstrate an improvement of more than 30% in logical generalization over baselines while preserving editing reliability and specificity. We further address evaluation biases in existing benchmarks through knowledge-aware protocols that disentangle external dependencies. This work establishes new state-of-the-art performance on ripple effect while ensuring internal logical consistency after knowledge editing.

Figures

Figures reproduced from arXiv: 2507.08427 by the authors.

Figure 1
Figure 1. After editing the president of USA, LLMs still [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. ChainEdit consists of three phases: (1) mine logical rule from KG and then align with LLM; (2) process [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of MEMIT-Merge method results [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of MEMIT method results on the original dataset and our variant dataset after using ChainEdit. original question”. Unlike previous evaluations re￾lying on internal knowledge integration, this setup tests LLMs’ capacity to combine edited knowledge with extern…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 6 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

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

    2024

  4. [4]

    Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2024. https://doi.org/10.1162/TACL\_A\_00644 Evaluating the ripple effects of knowledge editing in language models . Trans. Assoc. Comput. Linguistics, 12:283--298

    doi:10.1162/tacl 2024

  5. [5]

    Zilu Dong, Xiangqing Shen, and Rui Xia. 2025. https://doi.org/10.48550/arXiv.2502.07322 MEMIT - Merge : Addressing MEMIT 's Key - Value Conflicts in Same - Subject Batch Editing for LLMs . arXiv preprint. ArXiv:2502.07322 [cs]

    arXiv doi:10.48550/arxiv.2502.07322 2025

  6. [6]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lora: Low-rank adaptation of large language models . In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net

    2022

  7. [7]

    Jie Huang and Kevin Chen - Chuan Chang. 2023. https://doi.org/10.18653/V1/2023.FINDINGS-ACL.67 Towards reasoning in large language models: A survey . In Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023 , pages 1049--1065. Association for Computational Linguistics

    doi:10.18653/v1/2023.findings-acl.67 2023

  8. [8]

    Jiateng Liu, Pengfei Yu, Yuji Zhang, Sha Li, Zixuan Zhang, Ruhi Sarikaya, Kevin Small, and Heng Ji. 2024. https://aclanthology.org/2024.emnlp-main.282 EVEDIT: event-based knowledge editing for deterministic knowledge propagation . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, Novembe...

    work page 2024

Show all 20 references
  1. [9]

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

    2022

  2. [10]

    Andonian, Yonatan Belinkov, and David Bau

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

    2023

  3. [11]

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

    2022

  4. [12]

    Jiaxin Qin, Zixuan Zhang, Chi Han, Pengfei Yu, Manling Li, and Heng Ji. 2024. https://aclanthology.org/2024.emnlp-main.700 Why does new knowledge create messy ripple effects in llms? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMN...

    2024

  5. [13]

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

    2024

  6. [14]

    Jianchen Wang, Zhouhong Gu, Zhuozhi Xiong, Hongwei Feng, and Yanghua Xiao. 2024 a . https://doi.org/10.48550/ARXIV.2403.07825 The missing piece in model editing: A deep dive into the hidden damage brought by model editing . CoRR, abs/2403.07825

    2024 doi arXiv

  7. [15]

    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

  8. [16]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng...

    2024 arXiv

  9. [17]

    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

  10. [18]

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

    2024 doi arXiv

  11. [19]

    Zihao Zhao, Yuchen Yang, Yijiang Li, and Yinzhi Cao. 2024. https://aclanthology.org/2024.findings-emnlp.368 Ripplecot: Amplifying ripple effect of knowledge editing in language models via chain-of-thought in-context learning . In Findings of the Association for Computational L...

    2024

  12. [20]

    Manning, Christopher Potts, and Danqi Chen

    Zexuan Zhong, Zhengxuan Wu, Christopher D. Manning, Christopher Potts, and Danqi Chen. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.971 Mquake: Assessing knowledge editing in language models via multi-hop questions . In Proceedings of the 2023 Conference on Empirical Meth...

    2023 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.