Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read BalancEdit claims that giving every fact its own influence radius, stored in a detachable codebook, lets vision-language models be edited without retraining and without sacrificing generality or locality.

desk verdict A genuinely useful new benchmark and a plausible codebook-editing method, but the paper's central radius formula is written two opposite ways, so the balancing mechanism is irreproducible until that is fixed. read the letter →

arxiv 2505.01343 v2 pith:CP3RR4WV submitted 2025-05-02 cs.AI

Dongliang Guo , Mengxuan Hu , Zihan Guan , Thomas Hartvigsen , Sheng Li This is my paper · ORCID
classification cs.AI
keywords modeleditingmulti-modalgenerality-localitytrade-offdynamicinfluenceradiuscodebookOKEDITdatasetvisualquestionansweringsequential
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

BalancEdit sets out to show that a vision-language model can be updated fact by fact without retraining and without committing to one global influence scope. The paper's claim is that each edit should carry its own dynamic influence sphere, sized by a positive sample (a rephrased question) and a negative sample (a pure black image), and stored as an entry in a detachable codebook mounted on one transformer layer. Because the base weights never change, edits can be added, inspected, and reversed cheaply. The paper also introduces OKEDIT, a dataset built from OKVQA that measures text generality, image generality, and locality separately, and reports that BalancEdit reaches 100 percent editing success while keeping the generality-locality trade-off small. This matters because large multimodal models otherwise decay as facts change, and full retraining is too expensive for routine updates.

What carries the argument

The load-bearing object is the BalancEdit module, a discrete, localized codebook mounted on one transformer layer. Each entry is a triple $(k, v_k, \epsilon_k)$, and the lookup rule is $h^l_{i,t} = v_k(h^{l-1}_{i,t})$ when $\min d(h^{l-1}_{i,t}, K) \le \epsilon_k$, otherwise $f^l(h^{l-1}_{i,t})$. The radius is $\epsilon = (1-\alpha)\,d(\mathrm{Pos}, k) + \alpha\,d(\mathrm{Neg}, k)$, with Pos the rephrased question on the original image and Neg the original question on a pure black image; $\alpha$ controls how much weight the negative boundary gets. The radius is what makes the trade-off dynamic, because a broadly applicable fact receives a large sphere and a one-object fact a small one. The same rule also supports sequential edits, since later edits add keys and only reuse a prior transformation when their keys overlap.

What would settle it

For a single edit, take all images in a held-out category that is semantically related but out of scope (for example, other cat breeds when editing a breed name), compute their embeddings at the edited layer, and count how many fall within the stored radius; if a substantial fraction fall inside and trigger the edited transformation, the black-image anchor does not define the true influence boundary.

Watch

Extended reading notes

Core claim

The central claim is that the right response to the generality-locality trade-off is a per-fact, dynamically sized influence scope, not a fixed one. BalancEdit stores each edit as a codebook entry at a chosen transformer layer: a key $k$ is the averaged layer-$(l-1)$ embedding of the edit's image and question; a transformation $v_k$ is that layer's weights fine-tuned on the new answer; and an influence radius $\epsilon_k$ is set from positive and negative samples. At inference, the edited transformation applies only when the current embedding is within $\epsilon_k$ of the key; otherwise the original layer runs. The paper reports 100 percent editing success on both datasets with both backbones, sharp gains in text generality over the GRACE baseline, and the best harmonic mean of generality and locality among the baselines, with no pretraining and no data beyond the edit itself.

Load-bearing premise

The method rests on the assumptions that a pure black image is a universal stand-in for 'unrelated' inputs and that a rephrased question is always nearer to the edit key than any locality sample; if either fails, the radius will not trace the true boundary.

Editorial extensions

If this is right

  • A deployed vision-language model can be corrected one fact at a time in roughly eight seconds per edit, with no pretraining and no external data.
  • Every edit remains visible as a codebook entry, so an operator can inspect what was changed, reverse it, or remove it without touching the base model.
  • The same editing rule supports sequential updates: keys accumulate, and only overlapping or conflicting facts need to reuse or replace earlier transformations.
  • The $\alpha$ knob trades generality against locality, so a deployment can bias toward broad propagation or narrow, safe edits depending on the application.
  • Because the codebook is model-agnostic, the same edit mechanism can sit on different vision-language backbones by choosing the edited layer.

Reading between the lines

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

  • A natural next test is to replace the black-image anchor with a learned or sampled set of out-of-scope inputs, then compare the resulting radii against real proximities in the latent space.
  • The same positive/negative radius scheme transfers to text-only models by using a neutral or empty text prompt as the negative anchor, which would extend BalancEdit to language-model editing.
  • The reported locality drop under sequential edits suggests key collisions are the growth bottleneck, so a future version could merge or split entries when their radii overlap.
  • If distance functions or anchors are changed, the current results show the trade-off shifts, so a practical deployment might calibrate $\alpha$ per fact type rather than per model.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper addresses knowledge editing in large multimodal (vision-language) models. It formalizes a generality-locality trade-off, introduces an OKEDIT dataset built on OKVQA, and proposes BalancEdit, a memory-augmented editing method that stores per-edit codebook entries at a selected transformer layer. For each edit, BalancEdit constructs a positive sample by rephrasing the question and a negative sample using a black image, then sets an influence radius as a convex combination of the two embedding distances. The transformation is applied only to inputs whose embedding lies within the radius. Experiments compare BalancEdit with FT, IKE, MEND, and GRACE on two backbones (MiniGPT-4, BLIP-2 OPT) and two datasets (MMEDIT, OKEDIT), reporting high edit success and harmonic-mean trade-off scores. The paper claims no pretraining, interpretable codebooks, and dynamic per-edit balancing.

Significance. If the method works as claimed, BalancEdit would be a useful lightweight addition to the multimodal model editing toolkit: it is weight-preserving, requires no pretraining corpus, and its per-edit codebook gives interpretability and reversibility. The OKEDIT dataset is a potentially valuable resource for evaluating generality and locality, with more locality samples and harder semantic similarity than MMEDIT. The empirical HM improvements over MEND and GRACE on both backbones are non-trivial if they survive a corrected specification. Notable strengths are the released code and dataset, the absence of pretraining, and the interpretable codebook design. The main reservation is that the central radius mechanism is specified inconsistently in the paper, which currently prevents verification of the claimed dynamic balance; the theoretical appendix also falls short of establishing the mechanism.

major comments (3)
  1. [§3.2, Eq. (3); §4.8, Table 10; Appendix D, Eq. (10)] The paper gives two opposite conventions for the influence radius. Eq. (3) defines epsilon = (1-alpha)*d(Pos,k) + alpha*d(Neg,k), so alpha=0 gives the positive-sample distance and alpha=1 gives the negative-sample distance. In contrast, Section 4.8 and Table 10 state that alpha=0 makes the radius "entirely determined by the negative sample" (Loc=16.30) and alpha=1 makes it "determined solely by the positive sample" (I-Gen=24.22). Appendix D Eq. (10) also uses epsilon = alpha*S(f(G),f(E)) + (1-alpha)*S(f(L),f(E)), a convention opposite to Eq. (3). The Figure 5 narrative, in which increasing alpha shrinks the radius, is consistent only with the latter convention. Because alpha is the only parameter controlling the scope of each edit, a reader cannot determine which formula was implemented, and the headline claim of "dynamically balancing" generality and locality is not reproducible from the text.
  2. [Appendix D] The theoretical analysis is not a proof. The lemma that semantically similar concepts have close embeddings is asserted rather than derived, and the assumption S(G,E) < S(L,E) is exactly the separation condition needed for the conclusion. Eq. (10) then defines epsilon as an arbitrary convex combination of the two distances, which merely places epsilon between the two distances if the assumed ordering holds; it does not establish that this radius separates generality from locality, nor does it derive an optimal alpha. The phrase "we can find a radius" is vacuously true under the assumption and provides no support for the black-image negative anchor or for the claimed balance.
  3. [§3.2, "Influence Radius Determination"; Table 8] The assertion that a pure black image is a minimally informative, universal negative anchor is not validated. The distance from a black image to an edit key is not obviously related to the decision boundary between related and unrelated inputs for fact types as diverse as vehicles, people, plants, geography, and brands (the OKEDIT categories). Table 8 shows that using random negative samples changes I-Gen and Loc substantially, so the choice of anchor is consequential. The white-image ablation in Table 7 is only another null visual signal and does not validate the boundary property. The authors should provide per-category evidence that the black-image distance orders locality samples outside the radius, or weaken the claims about universal applicability.
minor comments (5)
  1. [Eq. (1); §4.8, Table 6] Eq. (1) uses a "<=" condition with a distance function, but Table 6 reports "cosine similarity" as a distance; for cosine, closer points have larger similarity, so the activation condition should be expressed via a proper distance transform such as 1-cosine.
  2. [§4.3] The text reports the selected alpha value for MiniGPT-4 (alpha=0.2) but not the corresponding value for BLIP-2 OPT; please report both selected values and the held-out set used.
  3. [Table 4] The "Sequential" column is ambiguous: the caption says it indicates whether the method uses sequential editing, but the baseline rows are marked with "✗" without explaining whether those baselines were evaluated in a non-sequential mode or cannot handle sequential edits at all.
  4. [Appendix A, Appendix D, §4.8, §4.3] There are several typos and formatting slips: "Fisrt" (Appendix A), "Theoratical Analyse" (Appendix D), "conducte" (Section 4.8), and "BLIP-2 OPT moder" (Section 4.3) should be corrected.
  5. [§4.4] The claim that locality improves by "20% to 80%" is imprecise because the comparison baseline is not fixed; please state the exact comparisons and values.

Circularity Check

2 steps flagged · score 6.0 of 10

Appendix D's radius 'proof' is self-definitional: Eq. (10) defines ε as a weighted average of the two distances, so Eq. (9) is true by construction; the α conventions in Eq. (3)/Fig. 2 and Sec. 4.8/Table 10 also contradict each other, making the balancing mechanism irreproducible.

  1. self definitional [Appendix D, Eqs. (9)-(10)]
    "Assumption: The generality sample (G) is semantically more similar to the editing knowledge (E) than the locality sample (L). That is, S(G,E ) < S(L,E ). ... Conclusion: In this case, we can find a radius ϵ such that S(f(G),f (E))<ϵ<S (f(L),f (E)), (9) where ϵ =α·S(f(G),f (E)) + (1−α)·S(f(L),f (E)). (10)"

    The conclusion is not a theorem about the model; it is the definition of ε. Once the ordering assumption S(f(G),f(E)) < S(f(L),f(E)) is granted, every convex combination ε = α·S(G) + (1−α)·S(L) with α in (0,1) automatically satisfies Eq. (9). Thus the claimed ability of BalancEdit's radius to separate generality from locality is baked into the construction rather than independently derived; the 'proof of effectiveness' reduces to the formula used to define the radius.

  2. other [Section 3.2, Eq. (3) and Fig. 2 caption vs. Section 4.8, Table 10 and Appendix D, Eq. (10)]
    "ϵ = (1−α)·d(Pos,k) +α·d(Neg,k), (3) ... When α = 0, the influence radius is entirely determined by the negative sample, which leads to over-generalization and reduced locality ... In contrast, when α = 1, the radius is determined solely by the positive sample, resulting in over-localization and poor generalization."

    This is a direct contradiction in the only mechanism that sets the influence scope. Eq. (3) and Fig. 2 make α=1 the negative-sample (large) radius, while Sec. 4.8/Table 10 and Appendix Eq. (10) make α=1 the positive-sample (small) radius. Since α is the sole control of the generality-locality balance and is tuned per model, the paper does not specify which construction was implemented. The headline 'dynamic balance' claim is therefore not reproducible or falsifiable from the text, compounding the self-definitional proof in Appendix D.

full rationale

The empirical part of the paper is largely self-contained: BalancEdit is compared against FT, IKE, MEND, and GRACE on MMEDIT and the new OKEDIT with defined metrics, ablations, and timing measurements, so the reported SOTA HM values are measurements rather than consequences of a fitted formula. The circularity is concentrated in the theoretical justification. Appendix D defines ε as a convex combination of the generality and locality distances after assuming the generality sample is closer; Eq. (9)'s separation inequality then follows by algebra, so the 'proof' is equivalent to its own definition. In addition, the paper gives two opposite conventions for α (Eq. 3/Fig. 2 vs. Sec. 4.8/Table 10/Eq. 10), so the sign of the balancing knob is ambiguous and the central mechanism cannot be reconstructed from the text. This is a specification defect rather than a circular reduction, but it bears directly on the paper's strongest claim. Self-citations, such as GRACE (Hartvigsen et al., 2024), are used as baselines or in the limitations discussion and are not load-bearing for BalancEdit's own derivation. Therefore the central theoretical prediction reduces by construction, while the empirical comparisons remain independent: partial circularity, score 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method's balancing behavior rests on two hand-designed anchors (rephrased question, black image) and a tuned interpolation coefficient alpha. No new physical or model-external entities are introduced; the codebook, keys, transformations, and radii are components of the proposed algorithm, not separate postulates.

free parameters (3)
  • alpha (radius interpolation coefficient) = 0.2 for MiniGPT-4; value for BLIP-2 OPT not stated
    Section 4.3: selected on a held-out set of 5 samples per model; controls the radius interpolation (Eq. 3) and thus the generality-locality balance. The reported HM scores are a function of this choice.
  • edited layer index = layer 31 mlp up_proj (MiniGPT-4), layer 31 fc2 (BLIP-2 OPT)
    Section 4.3: 'the choice of layer is a practical hyperparameter'; Table 9 shows HM varies with layer, so results are sensitive to this choice.
  • learning rate for transformation fine-tuning = 1e-2
    Section 4.3: best among 1.0, 1e-1, 1e-2, 1e-3, 1e-4, 1e-5; used for FT, MEND, GRACE, and BalancEdit's transformation updates.
assumptions (5)
  • domain assumption Semantically similar concepts have close embeddings in the model's latent space.
    Appendix D Lemma; used to justify approximating the influence scope by distances to positive/negative samples. This is a general empirical property, not established for every layer or fact type in this paper.
  • domain assumption A rephrased question with the same image is a valid positive sample representing the generality scope of an edit.
    Section 3.2 'Influence Radius Determination': 'rephrasing the text will not affect the semantic information of the edited knowledge.' Assumes LLM rephrasing preserves the edit-relevant semantics.
  • ad hoc to paper A pure black image is a universal negative anchor whose distance bounds the locality boundary.
    Section 3.2: 'The choice of black images as a proxy for out-of-scope knowledge is based on their characteristic as minimal or null visual signals.' No empirical validation across fact categories.
  • ad hoc to paper Generality samples are semantically closer to the edit than locality samples.
    Appendix D, Equation 8 and the surrounding assumption; the radius argument requires this ordering for every edit, but it is asserted rather than measured.
  • domain assumption Fine-tuning a single layer on one input reliably memorizes the new fact in the transformation v.
    Section 3.2 'Editing Transformations', Eq. 2: assumes the next-token prediction loss on one example yields a transformation that gives the target answer for the stored key.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing." pith.science (2026). https://pith.science/paper/CP3RR4WV

@misc{pith2026250501343,
  author       = {Pith},
  title        = {Pith review of: BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CP3RR4WV}},
  note         = {Machine review of arXiv:2505.01343}
}
read the original abstract

Large multi-modal models inevitably decay over time as facts update and previously learned information becomes outdated. Traditional approaches such as fine-tuning are often impractical for updating these models due to their size and complexity. Instead, direct knowledge editing within the models presents a more viable solution. Current model editing techniques, however, typically overlook the unique influence ranges of different facts, leading to compromised model performance in terms of both generality and locality. To address this issue, we introduce the concept of the generality-locality trade-off in multi-modal model editing. We develop a new model editing dataset named OKEDIT, specifically designed to effectively evaluate this trade-off. Building on this foundation, we propose \textbf{BalancEdit}, a novel method for balanced model editing that dynamically achieves an optimal balance between generality and locality. BalancEdit utilizes a unique mechanism that generates both positive and negative samples for each fact to accurately determine its influence scope and incorporates these insights into the model's latent space using a discrete, localized codebook of edits, without modifying the underlying model weights. To our knowledge, this is the first approach explicitly addressing the generality-locality trade-off in multi-modal model editing. Our comprehensive results confirm the effectiveness of BalancEdit, demonstrating minimal trade-offs while maintaining robust editing capabilities. Our code and dataset are available at https://github.com/donglgcn/BalancEdit/tree/MMOKVQA.

Figures

Figures reproduced from arXiv: 2505.01343 by the authors.

Figure 1
Figure 1. Illustration of various influence scope ods treat all influence scopes as if they are large and uni￾form, while others focus solely on a specific edit. For in￾stance, IKE (Zheng et al., 2023) employs in-context learn￾ing to edit knowledge, using the closest piece of knowl￾edge as a prompt to guide the language model. This ap￾proach causes the language model to rephrase the nearest fact, resulting in an oversized inf… view at source ↗
Figure 2
Figure 2. Overview of our BalancEdit framework. BalancEdit makes edits by learning, saving, and retrieving transformational edits between layers. The BalancEdit module consists of discrete keys, transformations, and a dynamic influence radius. Additionally, the BalancEdit module can handle multiple edits over time by adding new entries to the module. The formulation is as follows: h l i,t = ( vk(h l−1 i,t ), if min(d(h l−1 i,… view at source ↗
Figure 3
Figure 3. Illustration of influence radius determination without altering the overall architecture of the model. The parameters that are tuned include all the weights within the specific layer of the network. Influence Radius Determination. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: An example of the interpretable output training a meta-net. For instance, MEND, a meta-learning method, requires 22 hours to pre-train on 6,346 training samples. IKE, a retrieval-augmented in-context learning method, needs 12 hours to index 6,346 knowledge facts in adv…
Figure 5
Figure 5. Figure 5: Results of the effect of the hyperparameter α. typically update the knowledge within the model but do not provide a means to trace how these updates influence the model’s output. Specifically, while the model’s outputs may change, it is unclear how these changes are in…
Figure 6
Figure 6. Figure 6: Examples of our OKEDIT dataset. ‘CF’ represents the edited counterfactual knowledge. The red color indicates the out￾dated answer and the green color indicates the updated correct answer. the image should be ” After we obtain the image object, we can ask the diffusion …
Figure 7
Figure 7. Figure 7: T-sne figure of key distribution in sequential editing. by backpropagation with the target results. Keep handling the key conflicts could make it successfully deal with the multiple sequential editing in language models. However, to adapt it to the vision language mode…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Evaluating and Understanding Model Editing for Medical Vision Language Models

    cs.AI 2026-07 conditional novelty 6.0 of 10

    M3Bench is a clinically grounded benchmark showing that gradient-based VLM editors generalize but break locality, while memory-based editors preserve locality but fail on composition and temporal tasks, with failures ...

Reference graph

Works this paper leans on

48 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    L., and Parikh, D

    Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C. L., and Parikh, D. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pp.\ 2425--2433, 2015

    2015

  2. [2]

    Can we edit multimodal large language models? arXiv preprint arXiv:2310.08475, 2023

    Cheng, S., Tian, B., Liu, Q., Chen, X., Wang, Y., Chen, H., and Zhang, N. Can we edit multimodal large language models? arXiv preprint arXiv:2310.08475, 2023

    arXiv 2023

  3. [3]

    E., et al

    Chiang, W.-L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J. E., et al. Vicuna: An open-source chatbot impressing gpt-4 with 90\ See https://vicuna. lmsys. org (accessed 14 April 2023), 2 0 (3): 0 6, 2023

    2023

  4. [4]

    Dai, W., Li, J., Li, D., Tiong, A. M. H., Zhao, J., Wang, W., Li, B., Fung, P. N., and Hoi, S. Instructblip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  5. [5]

    Editing factual knowledge in language models

    De Cao, N., Aziz, W., and Titov, I. Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164, 2021

    arXiv 2021

  6. [6]

    Diederik, P. K. Adam: A method for stochastic optimization. 2014

    work page 2014

  7. [7]

    Making the v in vqa matter: Elevating the role of image understanding in visual question answering

    Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., and Parikh, D. 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, pp.\ 6904--6913, 2017

    2017

  8. [8]

    Guan, Z., Hu, M., Li, S., and Vullikanti, A. K. Ufid: A unified framework for black-box input-level backdoor detection on diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 27312--27320, 2025 a

    work page 2025

Show all 48 references
  1. [9]

    Benign samples matter! fine-tuning on outlier benign samples severely breaks safety

    Guan, Z., Hu, M., Zhu, R., Li, S., and Vullikanti, A. Benign samples matter! fine-tuning on outlier benign samples severely breaks safety. arXiv preprint arXiv:2505.06843, 2025 b

    2025 arXiv

  2. [10]

    Ada-VAD: Domain Adaptable Video Anomaly Detection, pp.\ 634--642

    Guo, D., Fu, Y., and Li, S. Ada-VAD: Domain Adaptable Video Anomaly Detection, pp.\ 634--642. doi:10.1137/1.9781611978032.73. URL https://epubs.siam.org/doi/abs/10.1137/1.9781611978032.73

    doi

  3. [11]

    Backdoor in seconds: Unlocking vulnerabilities in large pre-trained models via model editing, 2024

    Guo, D., Hu, M., Guan, Z., Guo, J., Hartvigsen, T., and Li, S. Backdoor in seconds: Unlocking vulnerabilities in large pre-trained models via model editing, 2024. URL https://arxiv.org/abs/2410.18267

    2024 arXiv

  4. [12]

    Aging with grace: Lifelong model editing with discrete key-value adaptors

    Hartvigsen, T., Sankaranarayanan, S., Palangi, H., Kim, Y., and Ghassemi, M. Aging with grace: Lifelong model editing with discrete key-value adaptors. Advances in Neural Information Processing Systems, 36, 2024

    2024

  5. [13]

    No free lunch: Retrieval-augmented generation undermines fairness in llms, even for vigilant users, 2024

    Hu, M., Wu, H., Guan, Z., Zhu, R., Guo, D., Qi, D., and Li, S. No free lunch: Retrieval-augmented generation undermines fairness in llms, even for vigilant users, 2024. URL https://arxiv.org/abs/2410.07589

    2024 arXiv

  6. [14]

    K., and Li, S

    Hu, M., Guan, Z., Zeng, Y., Guo, J., Zhou, Z., Zhang, J., Jia, R., Vullikanti, A. K., and Li, S. Mind control through causal inference: Predicting clean images from poisoned data. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  7. [15]

    Transformer-patcher: One mistake worth one neuron

    Huang, Z., Shen, Y., Zhang, X., Zhou, J., Rong, W., and Xiong, Z. Transformer-patcher: One mistake worth one neuron. arXiv preprint arXiv:2301.09785, 2023

    2023 arXiv

  8. [16]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Li, J., Li, D., Xiong, C., and Hoi, S. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pp.\ 12888--12900. PMLR, 2022

    2022

  9. [17]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Li, J., Li, D., Savarese, S., and Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pp.\ 19730--19742. PMLR, 2023

    2023

  10. [18]

    Pmet: Precise model editing in a transformer

    Li, X., Li, S., Song, S., Yang, J., Ma, J., and Yu, J. Pmet: Precise model editing in a transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 18564--18572, 2024

    2024

  11. [19]

    Moe-llava: Mixture of experts for large vision-language models

    Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., and Yuan, L. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024

    2024 arXiv

  12. [20]

    Liu, H., Li, C., Wu, Q., and Lee, Y. J. Visual instruction tuning, 2023 a

    2023

  13. [21]

    What makes good in-context examples for GPT -3? In Agirre, E., Apidianaki, M., and Vuli \'c , I

    Liu, J., Shen, D., Zhang, Y., Dolan, B., Carin, L., and Chen, W. What makes good in-context examples for GPT -3? In Agirre, E., Apidianaki, M., and Vuli \'c , I. (eds.), Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integra...

    2022 doi

  14. [22]

    Clip-driven universal model for organ segmentation and tumor detection

    Liu, J., Zhang, Y., Chen, J.-N., Xiao, J., Lu, Y., A Landman, B., Yuan, Y., Yuille, A., Tang, Y., and Zhou, Z. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 21152--21164, 2023 b

    2023

  15. [23]

    Ok-vqa: A visual question answering benchmark requiring external knowledge

    Marino, K., Rastegari, M., Farhadi, A., and Mottaghi, R. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pp.\ 3195--3204, 2019

    2019

  16. [24]

    Locating and editing factual associations in gpt

    Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt. In Advances in Neural Information Processing Systems, 2022 a

    2022

  17. [25]

    S., Andonian, A., Belinkov, Y., and Bau, D

    Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., and Bau, D. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229, 2022 b

    2022 arXiv

  18. [26]

    and Vondrick, C

    Menon, S. and Vondrick, C. Visual classification via description from large language models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=jlAjNL8z5cs

    2023

  19. [27]

    Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. arXiv preprint arXiv:2110.11309, 2021

    2021 arXiv

  20. [28]

    Mitchell, E., Lin, C., Bosselut, A., Finn, C., and Manning, C. D. Fast model editing at scale. In International Conference on Learning Representations, 2022 a

    2022

  21. [29]

    D., and Finn, C

    Mitchell, E., Lin, C., Bosselut, A., Manning, C. D., and Finn, C. Memory-based model editing at scale. In International Conference on Machine Learning. PMLR, 2022 b

    2022

  22. [30]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PMLR, 2021

    2021

  23. [31]

    A survey of hallucination in large foundation models

    Rawte, V., Sheth, A., and Das, A. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922, 2023

    2023 arXiv

  24. [32]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  25. [33]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

    Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., and Komatsuzaki, A. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021

    2021 arXiv

  26. [34]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35: 0...

    2022

  27. [35]

    Editable neural networks

    Sinitsin, A., Plokhotnyuk, V., Pyrkin, D., Popov, S., and Babenko, A. Editable neural networks. In International Conference on Learning Representations, 2020

    2020

  28. [36]

    Eva-clip: Improved training techniques for clip at scale

    Sun, Q., Fang, Y., Wu, L., Wang, X., and Cao, Y. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023

    2023 arXiv

  29. [37]

    Instructedit: Instruction-based knowledge editing for large language models

    Tian, B., Cheng, S., Liang, X., Zhang, N., Hu, Y., Xue, K., Gou, Y., Chen, X., and Chen, H. Instructedit: Instruction-based knowledge editing for large language models. arXiv preprint arXiv:2402.16123, 2024

    2024 arXiv

  30. [38]

    Visionllm: Large language model is also an open-ended decoder for vision-centric tasks

    Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. Advances in Neural Information Processing Systems, 36, 2024 a

    2024

  31. [39]

    Explainable anomaly detection in images and videos: A survey, 2024 b

    Wang, Y., Guo, D., Li, S., Camps, O., and Fu, Y. Explainable anomaly detection in images and videos: A survey, 2024 b . URL https://arxiv.org/abs/2302.06670

    2024 arXiv

  32. [40]

    Melo: Enhancing model editing with neuron-indexed dynamic lora

    Yu, L., Chen, Q., Zhou, J., and He, L. Melo: Enhancing model editing with neuron-indexed dynamic lora. arXiv preprint arXiv:2312.11795, 2023 a

    2023 arXiv

  33. [41]

    H., Kolehmainen, J., Shivakumar, P

    Yu, Y., Yang, C.-H. H., Kolehmainen, J., Shivakumar, P. G., Gu, Y., Ren, S. R. R., Luo, Q., Gourav, A., Chen, I.-F., Liu, Y.-C., et al. Low-rank adaptation of large language model rescoring for parameter-efficient speech recognition. In 2023 IEEE Automatic Speech Recognition a...

    2023

  34. [42]

    V., Mihaylov, T., Ott, M., Shleifer, S., Shuster, K., Simig, D., Koura, P

    Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X. V., Mihaylov, T., Ott, M., Shleifer, S., Shuster, K., Simig, D., Koura, P. S., Sridhar, A., Wang, T., and Zettlemoyer, L. Opt: Open pre-trained transformer language models, 2022

    2022

  35. [43]

    Can we edit factual knowledge by in-context learning? arXiv preprint arXiv:2305.12740, 2023

    Zheng, C., Li, L., Dong, Q., Fan, Y., Wu, Z., Xu, J., and Chang, B. Can we edit factual knowledge by in-context learning? arXiv preprint arXiv:2305.12740, 2023

    2023 arXiv

  36. [44]

    Panda: Prompt transfer meets knowledge distillation for efficient model adaptation

    Zhong, Q., Ding, L., Liu, J., Du, B., and Tao, D. Panda: Prompt transfer meets knowledge distillation for efficient model adaptation. arXiv preprint arXiv:2208.10160, 2022

    2022 arXiv

  37. [45]

    Navgpt: Explicit reasoning in vision-and-language navigation with large language models

    Zhou, G., Hong, Y., and Wu, Q. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 7641--7649, 2024

    2024

  38. [46]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models

    Zhu, D., Chen, J., Shen, X., Li, X., and Elhoseiny, M. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023

    2023 arXiv

  39. [47]

    A survey of trustworthy representation learning across domains

    Zhu, R., Guo, D., Qi, D., Chu, Z., Yu, X., and Li, S. A survey of trustworthy representation learning across domains. ACM Trans. Knowl. Discov. Data, 18 0 (7), June 2024. ISSN 1556-4681. doi:10.1145/3657301. URL https://doi.org/10.1145/3657301

    2024 doi

  40. [48]

    write newline

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

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.