Decoupling Reasoning and Knowledge Injection for In-Context Knowledge Editing

Changyue Wang; Qingyao Ai; Weihang Su; Yiqun Liu; Yujia Zhou

REVIEW 5 major objections 4 minor 3 cited by

Decoupling Reasoning and Knowledge Injection for In-Context Knowledge Editing

T0 review · 5 major / 4 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 In-context knowledge editing fails because injected facts derail the model's own reasoning; DecKER plans a masked reasoning skeleton before any edit is seen, then fills the slots, reaching 57.8% where the best baseline scores 54.3%.

desk verdict DecKER's plan-then-fill design is a genuine step forward for in-context knowledge editing, but its reasoning-preservation claim is partly tautological and the method's core assumption about fixed reasoning scaffolds is under-tested. read the letter →

arxiv 2506.00536 v1 pith:2KQFQMGS submitted 2025-05-31 cs.CL cs.AI

Changyue Wang , Weihang Su , Qingyao Ai , Yujia Zhou , Yiqun Liu This is my paper

classification cs.CLcs.AI

keywords knowledgeeditingin-contextmulti-hopquestionansweringreasoningpathplanninginjectionchain-of-thoughtconflictdetectionmasked

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 claims that in-context knowledge editing — updating an LLM's facts by adding new knowledge to the prompt rather than changing its weights — fails in multi-hop question answering because the injected facts entangle with the model's internal knowledge and derail its chain of reasoning. The authors' preliminary experiments support this: every tested method drops sharply after editing, and the drop tracks how much the post-editing reasoning path diverges from the pre-editing one. Their remedy, DecKER, generates a masked reasoning skeleton with placeholders and entity-type hints before any edit is introduced, then fills each placeholder step by step using retrieval from the edited memory, score-and-LLM conflict detection, and type-checked selection. Across three LLMs and four multi-hop QA benchmarks, DecKER-Base and its best-of-N variant outperform prior in-context and graph-based methods on most settings (57.8% vs. 54.3% for the strongest baseline RAE on MQuAKE-CF-3k-v2; 46.6% vs. 31.6% for EditCoT on RippleEdits-Popular). If correct, the work reframes editing as preserving the reasoning skeleton while replacing only the knowledge slots, a principle that would matter for any task where external content must coexist with parametric memory; the paper also reports that even DecKER loses some ground after editing (58.0% to 47.3% in its own preliminary set), so decoupling shrinks but does not erase the conflict.

What carries the argument

The load-bearing object is the masked reasoning path: a pre-edit skeleton in which every entity that knowledge must supply is replaced by a [MASK *] tag with a type hint, so the relation list $R_Q$ is fixed before any edited fact is seen. It is generated by a 5-shot prompt in which the model writes steps beginning with [STEP], numbers masks consistently, and declares entity categories; this skeleton is what stays identical before and after editing, guaranteeing the 1.00 reasoning-framework similarity. Filling is carried by two mechanisms: a hybrid conflict detector that declares a step edited when retrieval scores satisfy $S(d_1) > \alpha$ and $S(d_1) - S(d_2) > \beta$ (with an LLM check of whether the retrieved fact supports the masked sentence when the two criteria disagree), and a two-round selection pass — predictive entropy (Equation 4) retains the top half of sampled skeletons, and a yes/no entity-type check (Equation 5) picks the final filled path.

What would settle it

Partition the MQuAKE or RippleEdits questions by comparing the ground-truth relation list before and after editing (both derivable from the released reasoning paths and edit sets), and run DecKER separately on questions whose relation list is unchanged and on those where an edit changes the relations — for example, an edit that recategorizes an intermediate entity so the correct chain must now use a different relation. If the decoupling assumption holds, accuracy on the changed-relation questions should stay close to accuracy on the unchanged ones; a large drop on the changed-relation group would falsify the claim that the reasoning framework is independent of the edit set, and this comparison is directly reproducible from the paper's released data and code.

Watch

Extended reading notes

Core claim

The central claim is that a multi-hop question's reasoning framework — the ordered list of relations connecting the subject to the answer — should be treated as fixed before any edit is applied, and that the entanglement of reasoning and knowledge injection is what breaks existing in-context editing methods. The paper demonstrates this with a similarity metric comparing pre- and post-editing relation lists: MEMIT and the three ICE baselines score below 0.65, while DecKER scores 1.00 by construction, because the masked skeleton is generated from the question alone and the edited memory is consulted only during stepwise filling. At each masked step, a hybrid detector flags an edit when the top retrieval score exceeds $\alpha$ and the gap to the second result exceeds $\beta$, falling back to an LLM judgment when the two criteria disagree; non-edited slots are filled as fill-in-the-blank under a type hint. Multiple skeletons are sampled, the more confident half is kept via predictive entropy, and the final answer is chosen by entity-type match. The authors read the main results — DecKER-Base and DecKER-BoN ahead of all baselines on most of the twelve model-dataset cells — as confirming that preserving the reasoning framework is what drives the gains.

Load-bearing premise

The paper's method assumes that a question's reasoning plan — the ordered list of relations and entity types connecting question to answer — never changes when the facts are edited, so the masked skeleton generated before any edit is seen stays correct afterwards; an edit that changes an intermediate entity's type or the applicable relations makes the skeleton wrong, and the filling stage cannot repair it.

Editorial extensions

If this is right

DecKER-Base takes the top score on eleven of the twelve model-dataset cells in the main table, with EditCoT ahead only on MQuAKE-T for the two Qwen models; BoN reclaims the lead there for Qwen-14B.
In the batch-size study (1, 10, and 100 questions per editing batch on MQuAKE-300), DecKER-Base degrades more slowly than EditCoT as the batch grows, a direct consequence of decoupling: the planner never sees the edits, so batch growth only stresses the filling stage.
DecKER transfers to proprietary models: on MQuAKE-300 with GPT-4o-mini it scores 69.7% versus 67.0% for RAE, and the method itself needs no weight access or full output distributions, which is why the paper can run it on closed models.
Sampling helps: DecKER-BoN (N=6) improves over DecKER-Base on most settings by scoring sampled skeletons with predictive entropy plus entity-type fit (59.0 vs 57.8 for Llama-3.1-8B on MQuAKE-CF-3k-v2), while DecKER-Base is already the fastest strong method in the efficiency comparison.
The paper names fact verification, long-form generation, and retrieval-augmented generation as the next targets: RAG likewise injects retrieved content into the input and can clash with parametric knowledge, so the reasoning-first recipe is proposed as a fix there.

Reading between the lines

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

If the decoupling thesis is correct, the reasoning-framework similarity of Equation 2 is not just a diagnostic but a near-sufficient predictor: one could screen any editing method by measuring how much its relation list shifts after editing, without running full answer evaluation; the paper's own Figure 2 correlation is the seed of that predictor.
A testable boundary the paper exposes: edits that change an intermediate entity's type or the traversable relations break the pre-generated skeleton, and DecKER's own 58.0-to-47.3 post-edit drop is consistent with that residue. A direct extension is selective re-planning of the steps downstream of any fill that fails its type hint — the current method only filters such paths out at selection time
The same two-phase recipe should transfer to retrieval-augmented generation: plan the answer skeleton from the question alone, then treat each retrieved passage as a slot-filler. This is testable now, since the paper's masked-path prompts and hybrid conflict detector carry over almost unchanged, and it targets the known failure where retrieved context misleads the generator.
Because mask generation leans on the base model's planning ability, DecKER's ceiling is the LLM's own reasoning strength; the accuracy ordering across Qwen2.5-7B, Llama-3.1-8B, and Qwen2.5-14B is consistent with that, and it predicts that stronger reasoners should show larger absolute gains from decoupling, not just higher absolute scores.

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.

Desk Editor's Note

read the letter

DecKER is a real empirical contribution. The plan-then-fill architecture — generate a masked reasoning path before seeing any edits, then fill placeholders with retrieved or model-supplied entities — is genuinely new relative to Mello, PokeMQA, and EditCoT, which interleave planning with editing and get derailed by conflicting context. The evaluation is broad: three open-weight LLMs, four datasets, ablations, an efficiency comparison, and a code release. The main results are consistent, and the accuracy gains over baselines like RAE and EditCoT are large enough to take seriously. I would not be surprised if this becomes a reference point for in-context editing work.

The soft spots are real but addressable. Most important: the reasoning-framework similarity of 1.00 for DecKER is tautological — the same masked path is reused before and after editing, and the paper admits this in §4.4. That is a design property, not evidence that reasoning is preserved. The accuracy gains are not forced by this construction, so the core empirical claim holds, but the framing in §3.2 and Figure 2 overreaches. A fair reviewer should push for an evaluation of whether the filled path actually follows a sensible relation sequence, rather than similarity to itself.

Second, the method assumes the relation list in the reasoning skeleton remains valid after edits. If an edit changes an intermediate entity's type or makes a downstream relation inapplicable, the pre-generated skeleton is structurally wrong and the filling stage cannot repair it. The paper never tests this failure mode, and the benchmarks likely construct type-preserving edits, so the claimed advantage may only hold on that distribution. This should be stated as a limitation and ideally probed with a stress test.

Third, the GPT-4o-mini experiment is ambiguous: the text says the model is tested, but then says Llama-3.1-8B is used as a proxy for baselines that cannot run on proprietary APIs. That needs clarification. The hand-set hyperparameters (α, β, N) and the 300-question preliminary study without error bars are minor issues; the ablations partially address sensitivity, and the main results are consistent across settings.

This paper deserves a serious referee. The central idea is sound, the evidence is substantial, and the weaknesses are the kind a good revision can fix. I would read a revised version with interest.

Referee Report

5 major / 4 minor

Summary. The paper proposes DecKER, an in-context knowledge editing (ICE) method for multi-hop QA. The core idea is to decouple reasoning from knowledge injection: first generate a masked reasoning path (a relation-level skeleton with masked entities and type hints) before seeing any edits, then fill each masked entity step by step using a hybrid of retrieval-based conflict detection (score thresholds plus an LLM judgment) and LLM-based fill-in-the-blank. A best-of-N variant (DecKER-BoN) samples multiple masked paths and selects the final answer using predictive entropy and entity-type matching scores. Experiments on MQuAKE-CF-3k-v2, MQuAKE-T, RippleEdits-Popular, and RippleEdits-Random with Llama-3.1-8B, Qwen2.5-7B/14B, and GPT-4o-mini show consistent improvements over existing ICE methods (Mello, PokeMQA, DeepEdit, EditCoT, RAE) and over the parameterized baselines MEMIT and AlphaEdit.

Significance. If the reported accuracy gains hold, DecKER is a practical and conceptually interesting contribution to ICE: it does not require retraining, works on proprietary models, and is substantially faster than the strongest graph-based baseline RAE. The evaluation is broad (three open models plus GPT-4o-mini, four datasets, both base and BoN variants) and the code is released, which strengthens the reproducibility of the work. The main limitation is that the paper's central explanatory claim—that DecKER 'preserves the reasoning framework'—is partly true by construction, and the central assumption behind the decoupling (that the reasoning framework is independent of the edit set) is neither tested nor characterized. Those issues do not invalidate the empirical accuracy results, but they narrow the scope of what the paper can claim.

major comments (5)

[§4.4 and §6.1] The claim that DecKER 'maintains complete consistency in reasoning frameworks' is tautological: DecKER reuses the same masked reasoning path before and after editing, so the framework similarity reported as 1.000 in Figure 2 is 1 by construction. This cannot serve as empirical evidence that DecKER preserves reasoning integrity. The meaningful evidence is the post-edit accuracy drop (DecKER-Base falls from 58.0% to 47.3% in the same figure), which is not analyzed. The sentence 'This highlights the reason behind its strong performance' (§4.4) and the discussion in §6.1 that 'Preserving the reasoning framework is crucial' therefore conflate construction with evidence. I recommend reframing the reasoning-preservation claim as a design property, and reporting an analysis of the residual post-edit failures (e.g., how many come from retrieval misses, LLM fill errors, or format failures).
[§3.2 and §4.1] The paper's core assumption is that the reasoning framework R_Q (the ordered relation list) is independent of the edit set E, stated in §3.2 as 'the relation lists in the paths should remain unchanged after editing.' DecKER generates the masked skeleton before seeing edits (§4.1), and the filling stage (§4.2) only substitutes entities into existing relation slots. If an edit changes the entity type of an intermediate object (e.g., a person becomes a company) or makes a downstream relation inapplicable, the pre-generated skeleton is structurally wrong and cannot be repaired. This failure mode is not tested, is not mentioned in §8, and the benchmarks used (MQuAKE, RippleEdits) very likely preserve entity types and relation paths by construction. Since this assumption underlies the entire decoupling thesis, the paper should either provide evidence of relation-path invariance in the existing benchmarks (e.g., count how many edits preserve the relation sequence) or construct an experiment with type-changing edits to delimit the method's applicability.
[Table 1] The central claim in §6.1 is that DecKER 'outperforms previous methods across most models and datasets,' but no variance estimates or statistical tests are reported. Several margins are small, for example on MQuAKE-T with Llama-3.1-8B: DecKER-Base 80.9 versus PokeMQA 77.8 and DecKER-BoN 81.3 versus DecKER-Base 80.9. Without standard deviations across repeated runs, bootstrap confidence intervals, or paired tests, it is not possible to tell whether these differences reflect stable gains or sampling noise. Given the emphasis on 'significantly outperforms,' at least error bars over multiple seeds (or bootstrap resampling from the existing runs) should be provided.
[§3.2] The preliminary study that motivates the entire paper uses only 300 questions from MQuAKE-CF-3k-v2 and reports no error bars or repeated sampling. The claims that ICE can reduce accuracy 'by over 80%' and that editing performance correlates with reasoning-framework similarity are based on this small, single-run experiment. Since these results are used to justify the decoupling design, the paper should report confidence intervals (e.g., bootstrap over questions) or use a larger sample. This concern does not undermine the main results on the full datasets, but it weakens the paper's empirical motivation.
[§5.3, §6.3] The hyperparameters α=1.5, β=0.1, and N=6, as well as the sampling temperature 1.2 and top-p 0.95, are set in an ad hoc manner. The ablation in Figure 6 shows qualitative robustness of α and β within a range, but there is no systematic selection procedure, no error bars, and no analysis of how the optimal thresholds depend on the retriever or the corpus. Since retrieval score distributions vary across retrievers (contriever-msmarco here) and across edit sets, the transferability of α and β is not established. A small sensitivity table with confidence intervals, or a principled validation-set selection, would make the method more reproducible.

minor comments (4)

[Eq. (2)] The similarity measure pads the shorter relation list with zeros, but the underlying Sim function is a cosine similarity, which can be negative; the paper does not discuss how negative similarities are handled. This does not affect the main results but could be clarified.
[Figure 2] The caption refers to 'The second image,' which is the right panel, and the figure has no x-axis label on the right panel; using 'right panel' and adding an x-axis label (e.g., 'Method') would improve readability.
[Figure 4] The figure truncates the x-axis and the caption says so, but it is hard to locate specific data points; showing the full axis or using a log scale would help the reader compare average time per problem.
[§3.2] The comparison of reasoning-framework similarity uses GPT-4o-mini and jina-embeddings-v3, but the paper does not report any manual validation or inter-annotator agreement on the relation extraction; a small human-checked sample would increase confidence in the similarity numbers.

Circularity Check

1 steps flagged · score 4.0 of 10

Reasoning-framework consistency is by construction; the main accuracy comparison remains independent.

self definitional [Section 4.4 (Discussion), used with Eq. 2 in Section 3.2]
"DecKER maintains complete consistency in reasoning frameworks between pre- and post-editing, as they share the same masked reasoning path, highlighting the reason behind its strong performance."

The 'consistency in reasoning frameworks' is not an empirically measured outcome. Equation 2 defines reasoning-framework similarity as the similarity between the pre-edit and post-edit relation lists R1 and R2. In DecKER, the post-edit relation list is literally the same masked reasoning path generated before editing, with only entity slots filled in. Therefore the similarity is 1.0 by construction, regardless of whether the edited facts are injected correctly or whether the final answer is accurate. Using this score to 'highlight the reason behind its strong performance' makes the reasoning-preservation explanation tautological: the metric rewards the algorithm for reusing its own skeleton.

full rationale

The paper's headline accuracy claims are empirical and largely self-contained: DecKER-Base and DecKER-BoN are compared against external baselines on benchmark datasets, and the filling, conflict-detection, and answer-selection stages produce results that are not determined solely by the masked-path reuse. The circular component is narrower: the reasoning-framework-similarity result reported in Figure 2 and Section 4.4 is entailed by design, because the pre-edit and post-edit frameworks are the same masked skeleton. This makes the explanatory claim 'preserving the reasoning framework' definitional rather than evidential, but it does not reduce the main accuracy comparison to a fit or to a self-citation chain. The ablation of conflict-detection hyperparameters is performed on MQuAKE-300, but the paper does not explicitly state that alpha and beta were tuned on the test set, so I do not treat that as a fitted-input circularity. Accordingly, the score reflects one by-construction metric used as supporting evidence, while the central claim retains independent empirical content.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the LLM's ability to plan a relation skeleton without seeing edited facts, on the edit-independence of the relation skeleton, on a retrieval-based conflict heuristic introduced by the authors (α,β), and on reliable fill-in-the-blank generation for untouched entities. These are reasonable domain assumptions for multi-hop QA benchmarks but are not formally justified; the paper's ablations provide partial, qualitative support.

free parameters (4)

α (conflict detection high-score threshold) = 1.5
Used in Eq. 3a; set by hand; robustness shown only qualitatively in Figure 6.
β (top-2 retrieval score gap threshold) = 0.1
Used in Eq. 3b; set by hand; sensitivity analysis is qualitative.
N (number of sampled masked paths in DecKER-BoN) = 6
Chosen empirically; Appendix D shows performance plateaus around N=6.
Sampling temperature and top-p = 1.2, 0.95
Nucleus sampling parameters for path sampling; standard decoding choices, not derived from theory.

assumptions (5)

domain assumption The reasoning framework R_Q (ordered relation list) is independent of the edit set E; edits change objects but not the relation sequence.
Stated in §3.2 and used to justify generating the masked path before seeing edits; if false, the skeleton is invalid.
domain assumption The LLM can produce a correct relation skeleton for the masked reasoning path using only parametric knowledge and the question.
The entire method depends on this planning step; the paper notes in Limitations that DecKER relies solely on the LLM's own reasoning capabilities.
ad hoc to paper Retrieval score features (high top score, large top-2 gap) indicate an edited fact relevant to the masked step.
Eq. 3a/3b define the conflict detector; this heuristic is introduced by the paper and validated only through ablations.
domain assumption Each edited fact has a one-to-one mapping to an original fact, enabling conflict detection against the original.
Assumed in §4.2: 'Each edited fact maps to an original one'; required for the LLM conflict check.
domain assumption Fill-in-the-blank generation for non-edited entities is reliable when guided by type hints.
Used in §4.2 filling step; errors here propagate to the final answer, mitigated only by answer selection in BoN.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Decoupling Reasoning and Knowledge Injection for In-Context Knowledge Editing." pith.science (2026). https://pith.science/paper/2KQFQMGS

@misc{pith2026250600536,
  author       = {Pith},
  title        = {Pith review of: Decoupling Reasoning and Knowledge Injection for In-Context Knowledge Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KQFQMGS}},
  note         = {Machine review of arXiv:2506.00536}
}

read the original abstract

Knowledge editing aims to efficiently update Large Language Models (LLMs) by modifying specific knowledge without retraining the entire model. Among knowledge editing approaches, in-context editing (ICE) offers a lightweight solution by injecting new knowledge directly into the input context, leaving model parameters unchanged. However, existing ICE approaches do not explicitly separate the newly injected knowledge from the model's original reasoning process. This entanglement often results in conflicts between external updates and internal parametric knowledge, undermining the consistency and accuracy of the reasoning path.In this work, we conduct preliminary experiments to examine how parametric knowledge influences reasoning path planning. We find that the model's reasoning is tightly coupled with its internal knowledge, and that naively injecting new information without adapting the reasoning path often leads to performance degradation, particularly in multi-hop tasks. To this end, we propose DecKER, a novel ICE framework that decouples reasoning from knowledge editing by generating a masked reasoning path and then resolving knowledge edits via hybrid retrieval and model-based validation. Experiments on multi-hop QA benchmarks show that DecKER significantly outperforms existing ICE methods by mitigating knowledge conflicts and preserving reasoning consistency. Our code is available at: https://github.com/bebr2/DecKER .

Figures

Figures reproduced from arXiv: 2506.00536 by the authors.

**Figure 3.** The workflow of DecKER, only showing the processing of a single path, omitting the final answer [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 5.** #Hop vs. Performance: performance on the MQuAKE-CF-3k [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: Ablation studies for conflict detection on [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Edit Batch Size vs. Performance. The model [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

MetaSyn: A Benchmark for LLM Agents on Meta-Analysis Articles from Nature Portfolio
cs.CL 2026-06 unverdicted novelty 8.0 of 10

MetaSyn is a stage-level benchmark of 442 meta-analyses showing LLM agents retrieve up to 90.9% of eligible studies but include at most 52.7% in their final reports.
RWGBench: Evaluating Scholarly Positioning in Related Work Generation
cs.DL 2026-05 unverdicted novelty 7.0 of 10

RWGBench measures related-work generation by citation choices, and shows citation-focused metrics expose failures that text-similarity and LLM-judge scores miss.
Dynamic and Parametric Retrieval-Augmented Generation
cs.CL 2025-06 unverdicted novelty 2.0 of 10

A tutorial outline that categorizes recent RAG work into Dynamic RAG and Parametric RAG, and explains why both are needed.

Reference graph

Works this paper leans on

19 extracted references · 12 canonical work pages · cited by 3 Pith papers

[4]

In Proceedings of the ACM on Web Conference 2025, pages 4386–4395

Decoupling knowledge and context: An efficient and effective retrieval aug- mented generation framework via cross attention. In Proceedings of the ACM on Web Conference 2025, pages 4386–4395. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, An...

work page 2025
[6]

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov

Should we re- ally edit language models? on the evaluation of edited language models.Preprint, arXiv:2410.18785. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov

arXiv
[8]

InThe Eleventh International Conference on Learning Representa- tions, ICLR 2023, Kigali, Rwanda, May 1-5,

Mass- editing memory in a transformer. InThe Eleventh International Conference on Learning Representa- tions, ICLR 2023, Kigali, Rwanda, May 1-5,

work page 2023
[9]

InPro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM 2024, Boise, ID, USA, October 21-25, 2024, pages 2056–2066

Retrieval-enhanced knowledge editing in language models for multi-hop question answering. InPro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM 2024, Boise, ID, USA, October 21-25, 2024, pages 2056–2066. ACM. Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wa...

work page 2024
[10]

Weihang Su, Qingyao Ai, Xiangsheng Li, Jia Chen, Yiqun Liu, Xiaolong Wu, and Shengluan Hou

jina-embeddings- v3: Multilingual embeddings with task lora.Preprint, arXiv:2409.10173. Weihang Su, Qingyao Ai, Xiangsheng Li, Jia Chen, Yiqun Liu, Xiaolong Wu, and Shengluan Hou. 2024a. Wikiformer: Pre-training with structured information of wikipedia for ad-hoc retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages ...

arXiv 2024
[12]

Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu

Rbft: Robust fine-tuning for retrieval-augmented generation against retrieval de- fects.arXiv preprint arXiv:2501.18365. Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2024a. Knowledge editing through chain-of- thought.arXiv preprint arXiv:2412.17727. Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2024b. Knowledge editing through chain-of- t...

arXiv 2024
[14]

Chain-of-thought prompting elicits reasoning in large language models. InAd- vances in Neural Information Processing Systems 35: Annual Conference on Neural Information Process- ing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9,

work page 2022
[15]

InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online

Trans- formers: State-of-the-art natural language processing. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Wanli Yang, Fei Sun, Jiajun Tan, Xinyu Ma, Du Su, Dawei Yin, and Huawei Shen

work page 2020

Show all 19 references

[16]

InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 4079–4087

The fall of rome: Understanding the collapse of llms in model editing. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 4079–4087. Zexuan Zhong, Zhengxuan Wu, Christopher Manning, Christopher Potts, and Danqi Chen

2024
[17]

InProceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing, pages 15686–15702, Singapore

MQuAKE: Assessing knowledge editing in language models via multi-hop questions. InProceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing, pages 15686–15702, Singapore. Association for Computational Linguistics. A Pseudocode Description of DecK...

2023
[18]

The nullspace threshold for AlphaEdit is set to 0.02

and AlphaEdit (Fang et al., 2024a):We adopt the official im- plementation of AlphaEdit, setting the target edit- ing layers to 4-8, with a learning rate of 0.1 and a weight decay of 0.5. The nullspace threshold for AlphaEdit is set to 0.02. Mello (Zhong et al., 2023):We adhere...

2023
[19]

EditCoT (Wang et al., 2024b):We adhere to the original paper’s settings, conducting up to 4 rounds of CoT editing

RAE (Shi et al., 2024):We use the code pro- vided by the official implementation and adjust its retrieval process as described in subsection 5.4 to ensure a fair comparison. EditCoT (Wang et al., 2024b):We adhere to the original paper’s settings, conducting up to 4 rounds of C...

2024
[2018]

FEVER: a large-scale dataset for fact extraction and VERification. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809–819, New Orleans, Louisiana. A...

2018
[2020]

InAd- vances in Neural Information Processing Systems 33: Annual Conference on Neural Information Process- ing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

Language models are few-shot learners. InAd- vances in Neural Information Processing Systems 33: Annual Conference on Neural Information Process- ing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva

2020
[2021]

Yiwei Wang, Muhao Chen, Nanyun Peng, and Kai-Wei Chang

Kepler: A unified model for knowledge embedding and pre-trained language representation.Transac- tions of the Association for Computational Linguis- tics, 9:176–194. Yiwei Wang, Muhao Chen, Nanyun Peng, and Kai-Wei Chang. 2024d. Deepedit: Knowledge editing as de- coding with c...

arXiv
[2022]

Locating and editing factual associ- ations in GPT. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neu- ral Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - Decem- ber 9,

2022
[2023]

Giannis Bekoulis, Christina Papagiannopoulou, and Nikos Deligiannis

Longbench: A bilingual, multitask benchmark for long context understanding.arXiv preprint arXiv:2308.14508. Giannis Bekoulis, Christina Papagiannopoulou, and Nikos Deligiannis

arXiv
[2024]

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Xiang Wang, Xiangnan He, and Tat seng Chua

The llama 3 11 herd of models. Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Xiang Wang, Xiangnan He, and Tat seng Chua. 2024a. Alphaedit: Null-space constrained knowledge editing for language models.Preprint, arXiv:2410.02355. Yan Fang, Jingtao Zhan, Qingyao Ai, Jiaxin ...

arXiv
[2025]

Qian Dong, Qingyao Ai, Hongning Wang, Yiding Liu, Haitao Li, Weihang Su, Yiqun Liu, Tat-Seng Chua, and Shaoping Ma

Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforce- ment learning.Preprint, arXiv:2501.12948. Qian Dong, Qingyao Ai, Hongning Wang, Yiding Liu, Haitao Li, Weihang Su, Yiqun Liu, Tat-Seng Chua, and Shaoping Ma

arXiv

Pith tools

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

[1] [4]

In Proceedings of the ACM on Web Conference 2025, pages 4386–4395

Decoupling knowledge and context: An efficient and effective retrieval aug- mented generation framework via cross attention. In Proceedings of the ACM on Web Conference 2025, pages 4386–4395. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, An...

work page 2025

[2] [6]

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov

Should we re- ally edit language models? on the evaluation of edited language models.Preprint, arXiv:2410.18785. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov

arXiv

[3] [8]

InThe Eleventh International Conference on Learning Representa- tions, ICLR 2023, Kigali, Rwanda, May 1-5,

Mass- editing memory in a transformer. InThe Eleventh International Conference on Learning Representa- tions, ICLR 2023, Kigali, Rwanda, May 1-5,

work page 2023

[4] [9]

InPro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM 2024, Boise, ID, USA, October 21-25, 2024, pages 2056–2066

Retrieval-enhanced knowledge editing in language models for multi-hop question answering. InPro- ceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM 2024, Boise, ID, USA, October 21-25, 2024, pages 2056–2066. ACM. Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wa...

work page 2024

[5] [10]

Weihang Su, Qingyao Ai, Xiangsheng Li, Jia Chen, Yiqun Liu, Xiaolong Wu, and Shengluan Hou

jina-embeddings- v3: Multilingual embeddings with task lora.Preprint, arXiv:2409.10173. Weihang Su, Qingyao Ai, Xiangsheng Li, Jia Chen, Yiqun Liu, Xiaolong Wu, and Shengluan Hou. 2024a. Wikiformer: Pre-training with structured information of wikipedia for ad-hoc retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages ...

arXiv 2024

[6] [12]

Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu

Rbft: Robust fine-tuning for retrieval-augmented generation against retrieval de- fects.arXiv preprint arXiv:2501.18365. Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2024a. Knowledge editing through chain-of- thought.arXiv preprint arXiv:2412.17727. Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2024b. Knowledge editing through chain-of- t...

arXiv 2024

[7] [14]

Chain-of-thought prompting elicits reasoning in large language models. InAd- vances in Neural Information Processing Systems 35: Annual Conference on Neural Information Process- ing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9,

work page 2022

[8] [15]

InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online

Trans- formers: State-of-the-art natural language processing. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Wanli Yang, Fei Sun, Jiajun Tan, Xinyu Ma, Du Su, Dawei Yin, and Huawei Shen

work page 2020