Pith. sign in

REVIEW 4 major objections 5 minor 66 references

Towards Meta-Cognitive Knowledge Editing for Multimodal LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Multimodal knowledge editing needs a meta-cognitive layer, and the MIND framework supplies it.

desk verdict A creative benchmark and a plausible editor, but the meta-cognitive interpretation is not supported—two of three CogEdit levels can be scored by an unedited model. read the letter →

arxiv 2509.05714 v1 pith:YRTFHUDE submitted 2025-09-06 cs.AI cs.CV

Zhaoyu Fan , Kaihang Pan , Mingze Zhou , Bosheng Qin , Juncheng Li , Shengyu Zhang , Wenqiao Zhang , Siliang Tang
show 2 more authors
Fei Wu Yueting Zhuang
This is my paper
classification cs.AIcs.CV
keywords multimodalknowledgeeditingmeta-cognitionMLLMself-awarenessboundarymonitoringreflectivethinkingShapleyvaluebenchmark
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 argues that current multimodal knowledge editing only tests whether a model swaps old facts for new ones, missing whether the model knows when, why, and how to use the update. To close that gap, it builds CogEdit, a benchmark of 1,174 VQA-derived instances with counterfactual, boundary-constrained, and noisy-label editing conditions. It then proposes MIND, which adds a meta-knowledge memory, a Shapley-value-based activation monitor, and a prototype-based label refiner so that models can regulate their own knowledge updates. On MiniGPT-4 and LLaVA-1.5, MIND beats cognitive baselines on Adaptability, Compliance, and Clarity@K while staying competitive on the standard MMEdit benchmark. If the claim holds, editing an MLLM is not just about storing facts but about teaching the model self-aware, boundary-monitoring, and reflective behavior.

What carries the argument

The load-bearing object is MIND's meta-memory: a d'x d' projection matrix Mem applied to feed-forward activations, with each column a meta-knowledge unit. Units are split into meta-declarative knowledge, which records self-awareness of what the model knows, and meta-conditional knowledge, which records when to apply it. A Meta-memory Shapley Value (MSV) monitor, implemented as an MLP that approximates each unit's marginal contribution to editing performance, gates how strongly each memory unit activates. A prototype-based meta-label refiner then shifts the monitored activation toward retrieved label prototypes, with a mixing coefficient beta balancing raw and refined activations. CogEdit sup

What would settle it

Ask human raters to classify post-edit answers from MIND and SERAC on boundary-constrained items as showing genuine awareness of the condition or not; if MIND's higher Compliance does not predict higher human-rated awareness, the benchmark's construct validity fails.

Watch

Extended reading notes

Core claim

The central claim is that the standard Reliability, Generality, Locality trio does not capture a deeper layer of editing ability: a multimodal model can swap knowledge correctly and still fail to know when the new knowledge applies, when it should revert, and whether a noisy update is trustworthy. To measure those abilities, the authors construct CogEdit with three levels and new metrics, then claim that MIND achieves 'thinking about thinking' by writing into a meta-memory projection over feed-forward layers, gating activation through a Shapley-value-based monitor, and refining labels with a prototype bank. Across two architectures, MIND surpasses the cognitive baselines on Adaptability, Com

Load-bearing premise

CogEdit's Adaptability, Compliance, and Clarity@K scores are assumed to measure meta-cognition; if they actually measure ordinary robustness or instruction-following, the claim that MIND enables 'thinking about thinking' does not follow.

Editorial extensions

If this is right

  • A distinct meta-cognitive evaluation axis exists beyond Reliability, Generality, and Locality for judging edited multimodal models.
  • Editing methods that only store or patch facts, such as SERAC, IKE, MEND, WISE, and T-Patcher, could be augmented with a monitoring and refinement layer without discarding their parameter-preserving designs.
  • The meta-memory projection and MSV monitor apply to any transformer whose feed-forward layers are treated as key-value memories, not only the two MLLM families tested.
  • MIND's lifelong editing results suggest that a meta-cognitive memory can slow the degradation of editing quality as the number of sequential edits increases.
  • The paper's own limitations leave open whether these edits affect generative outputs and whether the mechanism scales to larger models.

Reading between the lines

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

  • CogEdit's three levels are behavioral tests; they do not by themselves establish that MIND has an internal reflective state, and adversarial paraphrasing could reveal pattern matching instead of true monitoring.
  • The MSV module is an MLP approximation of the Shapley value, so the 'game-theoretic' contribution is best read as learned importance gating rather than exact Shapley weighting.
  • The construct validity of Adaptability, Compliance, and Clarity@K is not externally validated; a human-rating protocol comparing MIND and SERAC answers would settle whether the metrics track meta-cognition.
  • A natural extension is to run MIND on language-only models and generative MLLMs; if the gains shrink, the mechanism may be tied to the visual-instruction setup rather than to meta-cognition itself.
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

4 major / 5 minor

Summary. The paper argues that existing multimodal knowledge editing operates only at a 'cognitive' level (reliably replacing old facts) and lacks 'meta-cognitive' abilities such as self-awareness, boundary monitoring, and reflective thinking. To support this, the authors introduce CogEdit, a benchmark with three levels—Counterfactual-Driven Editing, Boundary-Constraint Editing, and Noise-Robust Editing—measured by Fidelity/Adaptability, Reliability/Compliance, and Clarity@K, respectively. They also propose MIND, a framework that couples a meta-knowledge memory with a Shapley-value-based activation monitor and a prototype-based label refiner. Experiments on MiniGPT-4 and LLaVA report that MIND outperforms cognitive baselines on CogEdit and is competitive on the standard MMEdit benchmark, leading the authors to claim that MIND enables meta-cognitive knowledge editing.

Significance. If the CogEdit benchmark and the MIND results are valid, this would be a meaningful step toward measuring and implementing self-reflective knowledge updates in multimodal LLMs. The paper offers a concrete operationalization of meta-cognition, a new benchmark, and a method that is competitive on the established MMEdit task. The machine-checked aspects are limited, but the manuscript includes ablation studies, lifelong-editing experiments, and comparisons across two MLLM architectures. However, the central meta-cognitive claim rests almost entirely on CogEdit, whose construct validity is not established and whose Level 2 and Level 3 metrics may reward edit resistance rather than meta-cognitive processing. The Shapley-value monitoring component is also underspecified. The significance of the contribution is therefore contingent on addressing these validity concerns.

major comments (4)
  1. [Section 3 / Appendix A.2, Eqs. (9)-(11)] The CogEdit benchmark, as defined, does not rule out a no-op explanation. In Level 2, Reliability (Eq. 9) is evaluated on the original question with target ye equal to the original correct answer (e.g., 'Mercury'); an unedited model that already answers Mercury scores 100% without performing any edit. Compliance (Eq. 10) then asks a boundary-constrained question (e.g., 'Mars' under 'five farthest planets'); this tests pre-existing reasoning ability, not whether newly edited knowledge is applied within a learned boundary, since no knowledge change was actually introduced. In Level 3, Clarity@K (Eq. 11) removes the noisy label and asks for the original correct answer y; a model that refuses to incorporate the noisy edit can therefore score high. The paper does not report pre-edit accuracy or a no-edit baseline on CogEdit, so the observed MIND gains on Compliance and Clarity may reflect con
  2. [Section 4.2, Eqs. (3)-(5)] The MSV monitoring module is load-bearing for the 'game-theoretic monitoring' claim, but it is underspecified. Eq. (3) defines the Shapley value as a sum over subsets; Eq. (4) then writes phi_i = Softmax(Shapley(m_i, q)), which is neither the standard Shapley value nor derived from Eq. (3). The training procedure for the MLP monitor is not described: is the MLP trained with exact Shapley-value supervision? If so, how are exact values computed at the stated scale? If not, what loss is used? Eq. (5) also has a dimensional inconsistency: q_i is a scalar component of q, while m_i is a column of Mem; the product (q_i · m_i) and its elementwise combination with phi_i need precise definitions. Without these details, the 'Shapley-based monitoring' cannot be reproduced or verified.
  3. [Section 3, Data Collection] CogEdit is introduced as a measure of meta-cognitive ability, but no external validation is provided. The three levels and their metrics are asserted to correspond to self-awareness, boundary monitoring, and reflective thinking, yet no human study, comparison with established metacognition instruments, or analysis showing that the metrics diverge from generic reasoning/robustness is presented. Since the paper's central claim depends on CogEdit, the lack of construct validation is a substantive gap. At minimum, the authors should provide evidence that the metrics capture the intended constructs, e.g., by testing models with known meta-cognitive deficits or by obtaining human judgments of the benchmark instances.
  4. [Section 5, Tables 1-2 and Figure 5] No error bars, standard deviations, or statistical significance tests are reported for any experiment. Several headline differences are modest (e.g., LLaVA Compliance 59.08 vs. 53.20 for SERAC; Clarity@2 60.86 vs. 53.42). Without repeated runs or significance testing, the claim that MIND 'significantly outperforms' existing methods is not supported. This is particularly important because the proposed benchmark is new and the reported advantages are the main evidence for the meta-cognitive interpretation.
minor comments (5)
  1. [Eqs. (1)-(2)] The notation for the meta-memory projection is ambiguous: q_output = Mem(q_input) = [m_i(q_input_i)] with m_i a column vector and q_input_i a scalar. Please define the intended elementwise operation and the dimensions of all quantities.
  2. [References] The reference for Qizhou Chen et al. contains corrupted text: 'Life-long knowlecheng2023candge' should be cleaned and properly formatted.
  3. [Table 4 / Figure 3] The order of capability tiers in Table 4 is inconsistent with the order in the text and Figure 3 (Perception, Recognition, Understanding, Reasoning). Please align the presentation.
  4. [Appendix A.2] Formal task definitions and metrics for CogEdit appear only in the appendix. Since the benchmark is a core contribution, a concise formal summary should appear in the main text.
  5. [Section 5.2] The text states that MIND 'significantly' outperforms baselines, but no significance tests are presented; please soften the wording or add statistical support.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MIND's core claims are backed by an independent external benchmark (MMEdit), and the new CogEdit benchmark, while needing validity checks, is not derived from MIND's equations.

full rationale

The paper's derivation chain is not circular. MIND is a concrete architecture (meta-memory projection, MSV-based activation monitoring, prototype-based label refiner) and is evaluated on two benchmarks. CogEdit is a newly constructed benchmark, but constructing a benchmark and then optimizing it is not circular: the instances are curated from external VQA datasets with rule-based counterfactual/boundary/noise conditions, and the metrics are straightforward accuracy indicators. MMEdit provides an independent external benchmark where MIND is competitive, breaking any self-referential loop. The self-citations (Pan et al. 2025b for taxonomy/motivation, Lin et al. 2024 for MLP approximation of Shapley values) are not load-bearing: they support peripheral methodological choices, not the central claim. Concerns that CogEdit's Level 2 and Level 3 metrics may reward no-op behavior or that the metrics are not externally validated are construct-validity issues, not circularity of the kind this pass targets; they do not show that any equation or fitted parameter is equivalent to the claimed conclusion by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The central claim relies on several free parameters (the learned meta-memory, MSV monitor, prototype bank, projection matrix, and balance factor) that are fit to the editing data. The axioms include the key-value memory interpretation of FFNs, the author-defined three-level meta-cognitive framework, and the unvalidated assumption that CogEdit metrics measure meta-cognition. The invented entities are internal modules of MIND with no independent falsifiable handle outside the paper.

free parameters (5)
  • Meta-memory matrix Mem = Learned per edit, dimension d' x d'
    The core memory that stores meta-knowledge units; its entries are updated during editing to produce the desired output.
  • MSV monitor MLP = Learned weights, no details given
    Approximates Shapley values of meta-memory units; training data and loss are not specified.
  • Prototype bank P (K=64) = Learned embeddings
    Stores label prototypes used for label refinement; number of prototypes K is chosen as 64.
  • Label projection matrix Wp = Learned
    Projects retrieved prototype knowledge into a shifting vector in Eq. 6.
  • Balance factor beta = Not reported, presumably tuned
    Controls the trade-off between monitored activation and label shift in Eq. 6.
assumptions (5)
  • domain assumption Transformer feed-forward layers act as key-value memories (Geva et al., 2020)
    This justifies editing via feed-forward layer modification and is the foundation of the meta-memory design (Section 4.1).
  • ad hoc to paper Meta-cognition can be decomposed into self-awareness, boundary monitoring, and reflective thinking
    The three-level framework is introduced by the authors and used to define both the benchmark and the method; no external validation is provided.
  • domain assumption CogEdit metrics (Fidelity, Adaptability, Reliability, Compliance, Clarity@K) correctly measure the intended meta-cognitive capabilities
    The benchmark is constructed by the authors and not validated against human judgments or other benchmarks, yet the paper treats these metrics as ground truth.
  • ad hoc to paper The MSV monitor MLP accurately approximates the true Shapley values
    Section 4.2 claims the MLP is trained to predict Shapley values, but no training procedure, loss, or validation is described; the approximation is essential for the update rule in Eq. 5.
  • domain assumption Prototype-based label refinement improves noise robustness
    The label refiner is motivated by partial label learning (Wang et al., 2022) and is assumed to help extract useful information from noisy labels.
invented entities (3)
  • Meta-knowledge memory (meta-declarative and meta-conditional knowledge units)
    purpose: Enables self-aware knowledge updates by storing knowledge about knowledge
    Introduced as a new module in MIND; only evaluated indirectly through the CogEdit benchmark, no external evidence.
  • MSV-Based Activation Monitoring module
    purpose: Monitors knowledge activation via approximated Shapley values to guide meta-memory updates
    The MLP approximator is a new component whose accuracy is not independently validated.
  • Prototype-Based Meta-Label Refiner
    purpose: Filters noisy labels by retrieving and projecting label prototypes
    Adapts partial label learning ideas; its effectiveness is only shown in the paper's own ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Meta-Cognitive Knowledge Editing for Multimodal LLMs." pith.science (2026). https://pith.science/paper/YRTFHUDE

@misc{pith2026250905714,
  author       = {Pith},
  title        = {Pith review of: Towards Meta-Cognitive Knowledge Editing for Multimodal LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YRTFHUDE}},
  note         = {Machine review of arXiv:2509.05714}
}
read the original abstract

Knowledge editing enables multimodal large language models (MLLMs) to efficiently update outdated or incorrect information. However, existing benchmarks primarily emphasize cognitive-level modifications while lacking a focus on deeper meta-cognitive processes. To bridge this gap, we introduce CogEdit, a novel benchmark designed to evaluate MLLMs' meta-cognitive knowledge editing abilities across three levels: (1) Counterfactual-Driven Editing, assessing self-awareness of knowledge correctness changes; (2) Boundary Constraint Editing, ensuring appropriate generalization without unintended interference; and (3) Noise-Robust Editing, promoting reflective evaluation of uncertain information. To advance meta-cognitive editing, we propose MIND (Meta-cognitive INtegrated Dynamic Knowledge Editing), a framework that constructs a meta-knowledge memory for self-awareness, employs game-theoretic interactions to monitor knowledge activation, and incorporates label refinement for noise-robust updates. Extensive experiments show that MIND significantly outperforms existing cognitive editing approaches, achieving strong performance on both traditional and meta-cognitive knowledge editing benchmarks.

Figures

Figures reproduced from arXiv: 2509.05714 by the authors.

Figure 2
Figure 2. The performance of MIND and SERAC (Mitchell et al., 2022) on traditional cognitive editing benchmark and our proposed CogEdit. man et al., 2006). Meta-cognition encompasses a model’s ability to assess and modify its own rea￾soning and learning processes. In contrast to basic cognition, meta-cognition emphasizes the capacity to “think about thinking” and incorporates three additional levels. • Level 1: Self-Awareness… view at source ↗
Figure 3
Figure 3. (a) Distribution of problem difficulties in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. The MIND framework unifies meta-memory, monitoring mechanisms, and label refinement for meta [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Effectiveness of MSV-Monitored Update tive thinking: In noisy editing scenarios, MIND achieves higher Clarity than cognitive-based meth￾ods, highlighting its ability to analyze and extract useful information from noisy labels, demonstrat￾ing its reflective thinking abi…
Figure 6
Figure 6. Figure 6: Ablation study on meta-label refiner: (a) [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 22 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]

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425--2433

    2015

  4. [4]

    Sarah-Jayne Blakemore and Chris Frith. 2003. Self-awareness and action. Current opinion in neurobiology, 13(2):219--224

    work page 2003

  5. [5]

    John C Cavanaugh and Marion Perlmutter. 1982. Metamemory: A critical examination. Child development, pages 11--28

    work page 1982

  6. [6]

    Qizhou Chen, Chengyu Wang, Dakan Wang, Taolin Zhang, Wangyue Li, and Xiaofeng He. 2024 a . Lifelong knowlecheng2023candge editing for vision language models with low-rank mixture-of-experts. arXiv preprint arXiv:2411.15432

    arXiv 2024

  7. [7]

    Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Doll \'a r, and C Lawrence Zitnick. 2015. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325

    arXiv 2015

  8. [8]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2024 b . https://arxiv.org/abs/2312.14238 Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks . Preprint, arXiv:2312.14238

    arXiv 2024

Show all 66 references
  1. [9]

    Siyuan Cheng, Bozhong Tian, Qingbin Liu, Xi Chen, Yongheng Wang, Huajun Chen, and Ningyu Zhang. 2023 a . https://aclanthology.org/2023.emnlp-main.856 Can we edit multimodal large language models? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Pr...

    2023

  2. [10]

    Siyuan Cheng, Bozhong Tian, Qingbin Liu, Xi Chen, Yongheng Wang, Huajun Chen, and Ningyu Zhang. 2023 b . Can we edit multimodal large language models? arXiv preprint arXiv:2310.08475

    2023 arXiv

  3. [11]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. https://arxiv.org/abs/2305.06500 Instructblip: Towards general-purpose vision-language models with instruction tuning . Preprint, arXiv:2305.06500

    2023 arXiv

  4. [12]

    Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164

    2021 arXiv

  5. [13]

    Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. 2023. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010

    2023 arXiv

  6. [14]

    Petros Georghiades. 2004. From the general to the situated: Three decades of metacognition. International journal of science education, 26(3):365--383

    2004

  7. [15]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2020. Transformer feed-forward layers are key-value memories. arXiv preprint arXiv:2012.14913

    2020 arXiv

  8. [16]

    Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904--6913

    2017

  9. [17]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  10. [18]

    Tom Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2023. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems, 36:47934--47959

    2023

  11. [19]

    Han Huang, Haitian Zhong, Tao Yu, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. 2024. https://arxiv.org/abs/2403.07350 Vlkeb: A large vision-language model knowledge editing benchmark . Preprint, arXiv:2403.07350

    2024 arXiv

  12. [20]

    Zeyu Huang, Yikang Shen, Xiaofeng Zhang, Jie Zhou, Wenge Rong, and Zhang Xiong. 2023. Transformer-patcher: One mistake worth one neuron. arXiv preprint arXiv:2301.09785

    2023 arXiv

  13. [21]

    Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700--6709

    2019

  14. [22]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118

    2021 arXiv

  15. [23]

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/d89a66c7c80a29b1bdbab0f2a1a94af8-Paper.pdf Supervised contrastive learning . In A...

    2020

  16. [24]

    Emily R Lai. 2011. Metacognition: A literature review

    2011

  17. [25]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural informati...

    2020

  18. [26]

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. 2024 a . Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895

    2024 arXiv

  19. [27]

    Jiaqi Li, Miaozeng Du, Chuanyi Zhang, Yongrui Chen, Nan Hu, Guilin Qi, Haiyun Jiang, Siyuan Cheng, and Bozhong Tian. 2024 b . Mike: A new benchmark for fine-grained multimodal entity knowledge editing. arXiv preprint arXiv:2402.14835

    2024 arXiv

  20. [28]

    Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Wei Ji, Wenqiao Zhang, Tat-Seng Chua, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. 2023 a . Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. In The Twelfth International Conference on Learning Repre...

    2023

  21. [29]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023 b . https://arxiv.org/abs/2301.12597 Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models . Preprint, arXiv:2301.12597

    2023 arXiv

  22. [30]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023 c . Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730--19742. PMLR

    2023

  23. [31]

    Moxin Li, Yong Zhao, Yang Deng, Wenxuan Zhang, Shuaiyi Li, Wenya Xie, See-Kiong Ng, and Tat-Seng Chua. 2024 c . Knowledge boundary of large language models: A survey. arXiv preprint arXiv:2412.12472

    2024 arXiv

  24. [32]

    Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190

    2021 arXiv

  25. [33]

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

    2024

  26. [34]

    Tianwei Lin, Jiang Liu, Wenqiao Zhang, Zhaocheng Li, Yang Dai, Haoyuan Li, Zhelun Yu, Wanggui He, Juncheng Li, Hao Jiang, and 1 others. 2024. Teamlora: Boosting low-rank adaptation with expert collaboration and competition. arXiv preprint arXiv:2408.09856

    2024 arXiv

  27. [35]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning

    2023

  28. [36]

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2021. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602

    2021 arXiv

  29. [37]

    Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. 2019. https://arxiv.org/abs/1906.00067 Ok-vqa: A visual question answering benchmark requiring external knowledge . Preprint, arXiv:1906.00067

    2019 arXiv

  30. [38]

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

    2022

  31. [39]

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2022 b . Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229

    2022 arXiv

  32. [40]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2021. Fast model editing at scale. arXiv preprint arXiv:2110.11309

    2021 arXiv

  33. [41]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, and Chelsea Finn. 2022. Memory-based model editing at scale. In International Conference on Machine Learning, pages 15817--15831. PMLR

    2022

  34. [42]

    Thomas O Nelson. 1990. Metamemory: A theoretical framework and new findings. In Psychology of learning and motivation, volume 26, pages 125--173. Elsevier

    1990

  35. [43]

    Kaihang Pan, Wendong Bu, Yuruo Wu, Yang Wu, Kai Shen, Yunfei Li, Hang Zhao, Juncheng Li, Siliang Tang, and Yueting Zhuang. 2025 a . Focusdiff: Advancing fine-grained text-image alignment for autoregressive visual generation through rl. arXiv preprint arXiv:2506.05501

    2025 arXiv

  36. [44]

    Kaihang Pan, Zhaoyu Fan, Juncheng Li, Qifan Yu, Hao Fei, Siliang Tang, Richang Hong, Hanwang Zhang, and Qianru Sun. 2025 b . Towards unified multimodal editing with enhanced knowledge collaboration. Advances in Neural Information Processing Systems, 37:110290--110314

    2025

  37. [45]

    Kaihang Pan, Juncheng Li, Hongye Song, Jun Lin, Xiaozhong Liu, and Siliang Tang. 2023. Self-supervised meta-prompt learning with meta-gradient regularization for few-shot generalization. arXiv preprint arXiv:2303.12314

    2023 arXiv

  38. [46]

    Kaihang Pan, Juncheng Li, Wenjie Wang, Hao Fei, Hongye Song, Wei Ji, Jun Lin, Xiaozhong Liu, Tat-Seng Chua, and Siliang Tang. 2024 a . I3: I ntent-i ntrospective retrieval conditioned on i nstructions. In Proceedings of the 47th International ACM SIGIR Conference on Research a...

    2024

  39. [47]

    Kaihang Pan, Wang Lin, Zhongqi Yue, Tenglong Ao, Liyu Jia, Wei Zhao, Juncheng Li, Siliang Tang, and Hanwang Zhang. 2025 c . Generative multimodal pretraining with discrete diffusion timestep tokens. arXiv preprint arXiv:2504.14666

    2025 arXiv

  40. [48]

    Kaihang Pan, Siliang Tang, Juncheng Li, Zhaoyu Fan, Wei Chow, Shuicheng Yan, Tat-Seng Chua, Yueting Zhuang, and Hanwang Zhang. 2024 b . Auto-encoding morph-tokens for multimodal llm. arXiv preprint arXiv:2405.01926

    2024 arXiv

  41. [49]

    Kaihang Pan, Yang Wu, Wendong Bu, Kai Shen, Juncheng Li, Yingting Wang, Yunfei Li, Siliang Tang, Jun Xiao, Fei Wu, and 1 others. 2025 d . Unlocking aha moments via reinforcement learning: Advancing collaborative visual comprehension and generation. arXiv preprint arXiv:2506.01480

    2025

  42. [50]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. https://arxiv.org/abs/2307.01952 Sdxl: Improving latent diffusion models for high-resolution image synthesis . Preprint, arXiv:2307.01952

    2023 arXiv

  43. [51]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. 2022. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3

    2022 arXiv

  44. [52]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821--8831. Pmlr

    2021

  45. [53]

    Carol Rodgers. 2002. Defining reflection: Another look at john dewey and reflective thinking. Teachers college record, 104(4):842--866

    2002

  46. [54]

    Alvin E Roth. 1988. Introduction to the shapley value. The Shapley value, 1

    1988

  47. [55]

    Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. 2020. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827--6839

    2020

  48. [56]

    Marcel VJ Veenman, Bernadette HAM Van Hout-Wolters, and Peter Afflerbach. 2006. Metacognition and learning: Conceptual and methodological considerations. Metacognition and learning, 1:3--14

    2006

  49. [57]

    Haobo Wang, Ruixuan Xiao, Yixuan Li, Lei Feng, Gang Niu, Gang Chen, and Junbo Zhao. 2022. Pico+: Contrastive label disambiguation for robust partial label learning. arXiv preprint arXiv:2201.08984

    2022 arXiv

  50. [58]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, and 1 others. 2024 a . Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191

    2024 arXiv

  51. [59]

    Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. 2024 b . https://arxiv.org/abs/2405.14768 Wise: Rethinking the knowledge memory for lifelong model editing of large language models . Preprint, arXiv:2405.14768

    2024 arXiv

  52. [60]

    Renzhi Wang and Piji Li. 2024. Lemoe: Advanced mixture of experts adaptor for lifelong model editing of large language models. arXiv preprint arXiv:2406.20030

    2024 arXiv

  53. [61]

    Eyal Winter. 2002. The shapley value. Handbook of game theory with economic applications, 3:2025--2054

    2002

  54. [62]

    Junzhe Zhang, Huixuan Zhang, Xunjian Yin, Baizhou Huang, Xu Zhang, Xinyu Hu, and Xiaojun Wan. 2024 a . https://arxiv.org/abs/2406.13219 Mc-mke: A fine-grained multimodal knowledge editing benchmark emphasizing modality consistency . Preprint, arXiv:2406.13219

    2024 arXiv

  55. [63]

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

    2024 arXiv

  56. [64]

    Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023. Can we edit factual knowledge by in-context learning? arXiv preprint arXiv:2305.12740

    2023 arXiv

  57. [65]

    Zexuan Zhong, Zhengxuan Wu, Christopher D Manning, Christopher Potts, and Danqi Chen. 2023. Mquake: Assessing knowledge editing in language models via multi-hop questions. arXiv preprint arXiv:2305.14795

    2023 arXiv

  58. [66]

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.