- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposed DELTA, a multimodal knowledge-editing method that directly modifies the FFN output weights without additional training of an editor. It first selects the layers using the image-text pairs correlations, engagements and activation-score entropy. Then, DELTA ranks neurons using the modality-specific activation ratios and exclude the neurons heavily weighted by visual features. The overall method is optimized with prefix preservation, repetition, null-space regularization and visual perturbation constraints. Experiments with three main backbones and datasets plus additional Qwen results support the claim of this paper.
Critical strengths
- A good motivation from current unsolved gaps in the VLLM domain and concrete solutions to achieve the editing procedure. Multimodal editing expects to achieve the successful answers without hurting the visual and textual behaviours. DELTA, serves as a good method, balance the success and damage of VLLM editing task efficiently.
- The results reported in the paper strongly support the claims by evaluating among different backbones and datasets. DELTA has the highest reported average among that family in all nine main model–dataset settings. E.g. on LLaVA/VLKEB, its average is 94.74 while FiNE and ROME only have 87.45 and 83.28, respectively.
- The ablation experiments show useful evidence that the preservation components contribute. For example, on BLIP2/VLKEB, removing the null-space component reduces modal locality from 87.80 to 80.49; removing visual perturbation reduces it to 80.93. Reliability remains 100.00 in both cases.
Critical weaknesses
- The criteria to select the layers are not strong and robust. Activation correlation, mean cores and entropy are plausible but do not represent the causal fusion where the particular fact is included.
- The novelty claims need further improvements. Some prior work like M2Edit and NAM have done some overlapped jobs and DELTA may contribute a distinct localization rule and preservation strategy.
- Generality claims exceed the real capability of the paper. The localization formulation requires visual and textual positions to pass through common FFN neurons. But it does not directly cover a separate visual-memory architecture accessed through decoder cross-attention, which makes the paper architecture-specific.
Questions for authors
- In Appendix C, it constructs the activation matrix A from the edit inputs, controlling the direct FFN perturbation. However, it does not show the claim that the activation matrix building upon an unrelated input is small, which means the locality guarantee does not build the relationship to the unrelated inputs.
- A discrepancy occurs in Appendix I: Table 5 gives identical layer sets for configurations (a) and (e), while Table 7 reports average scores of 48.21 and 66.55. The paper does not explain them.
- What’s the significant improvements (theoretical and computational) from LoRA? Maybe more experiments of comparison between LoRA and DELTA are rational.
- Figure 1 has some overlapped arrows and labels which are hard to visualize. Figure 2, 3 and 6 are too small to see.
6 comments
Answer to Q1 and Q2
Q1. - Appendix C builds the activation matrix from the edit inputs only; it is not shown that the corresponding matrix of an unrelated input is small, so the locality guarantee has no relationship to unrelated inputs.
Thank you for your feedback. This is an important issue, and we would like to clarify. We can be candid about the incomplete theoretical analysis in the original version. That's why we supplement a more detailed theoretical justification for the null-space constraint (you can refer to our response to reviewer omH8) in Appendix D, providing two complementary views towards unrelated input preservation.
In this revision, we illustrate that the null-space construction is built from the edit input only, and the separation between unrelated activations and the edit targets is real exsisted: the locality bound on unrelated inputs holds only under a measurable separation condition (Assumption 1), and this condition is currently supported by our experiments (Appendix D.4, D.5). In Tab. 6 (also listed within the response to reviewer omH8), the unrelated-input next-token KL falls monotonically from to as reduces from to . The scale-invariant alignment falls from to . This matches Theorem 1's prediction that damage scales with and shrinks as the null cone tightens. We also measured Assumption 1 directly. % of unrelated positions have exactly zero trace-neuron activation.
These results show that part of the observed KL reduction on unrelated inputs comes from how shapes during optimization, rather than from strictly guaranteed orthogonality; the two panels of Tab. 6 separate the contributions: with the hard projection disabled, alone reduces the KL from to (), while the hard projection itself accounts for the near-removal (). We have therefore revised the claims in the paper, and we no longer claim an unconditional guarantee for arbitrary unrelated inputs. However, still empirically preserves unrelated knowledge to a meaningful extent and improves overall editing performance. We will keep on studying separation condition as our future works.
In addition, we also note that this preservation does not conflict with editability. Prompt positions are protected pointwise by Eqn. 40. Generation positions are by construction not rows of , and they retain null-cone mass. In our measurements, trace neurons are active at % of generation positions versus % of prompt positions. Generation positions carry % of the cone energy versus % for prompt positions. This is the editing channel required by Assumption 2, and it stays open. We have provided the more complete assumptions, lemmas, and proofs regarding this issue in Appendix D. We will appreciate it if you can spare more time on checking the appended theorems.
Q2. - Appendix I discrepancy: Table 5 gives identical layer sets for configurations (a) and (e), yet Table 7 reports average scores of 48.21 vs. 66.55, unexplained.
Thanks for the debugging! Reviewer HVFd also pointed out similar issues. This is caused by a statistical reporting error in the (d)-(f) row when we recomputed and typed the results into the paper. The intersection layers row (a)-(c) are computed based on the (d)-(f) rows, which are therefore wrongly calculated. The correct (a) and (e) configuration is as below:
| Notation | |
|---|---|
| (a) | |
| (e) |
, which should be different. We apologize for this oversight and recomputed the intersection layer set between each pair to fix Tab. 9. We once again checked the logs when running the experiments that produced Tab. 11 and confirmed the correctness of the runtime layer settings. We revised this reporting error in our revision and remarked the differences for review. Thanks.
Answer to Q3 and Q4
Q3. - What are the significant improvements (theoretical and computational) over LoRA? More LoRA-vs-DELTA comparison experiments would be rational.
We thank the reviewer for this insightful suggestion. Since DELTA directly updates the target neurons and the updating target can be decomposed into a matrix pair , LoRA can therefore be regarded as an extension of DELTA. We provide theoretical analysis based on the formulation in Appendix C and Appendix D. We conduct raw LoRA experiments based on the additional parameter-inserting layer given by VisEdit[3] with rank set to , compared to DELTA and the LoRA extension with for BLIP2 and for LLaVA, respectively. The results are as follows:
| Backbone | Method | #Edits=1 | #Edits=10 | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Rel | M-Gen | T-Gen | M-Loc | T-Loc | Avg | Rel | M-Gen | T-Gen | M-Loc | T-Loc | Avg | ||
| BLIP2-OPT-2.7B | LoRA (r=32) | 80.2 | 44.5 | 59.0 | 78.9 | 62.3 | 64.9 | 57.2 | 28.0 | 40.1 | 55.2 | 40.9 | 44.3 |
| DELTA (ours) | 96.3 | 94.9 | 85.5 | 84.0 | 99.6 | 92.0 | 86.0 | 76.1 | 76.9 | 61.6 | 91.9 | 78.5 | |
| +LoRA (r=4) | 98.5 | 97.6 | 93.1 | 59.7 | 97.3 | 89.3 | 79.5 | 75.6 | 69.9 | 28.4 | 89.9 | 68.7 | |
| +LoRA (r=8) | 100.0 | 99.4 | 95.1 | 57.1 | 97.0 | 89.7 | 78.7 | 76.9 | 71.7 | 27.5 | 91.1 | 69.2 | |
| +LoRA (r=32) | 100.0 | 99.5 | 95.1 | 52.8 | 96.9 | 88.9 | 79.4 | 75.0 | 71.7 | 24.8 | 89.9 | 68.1 | |
| LLaVA-1.5-7B | LoRA (r=32) | 88.3 | 45.3 | 64.2 | 39.5 | 82.4 | 63.9 | 42.8 | 39.6 | 50.7 | 54.4 | 30.1 | 43.5 |
| DELTA (ours) | 100.0 | 89.6 | 89.9 | 94.7 | 98.9 | 90.7 | 96.2 | 73.6 | 84.3 | 68.7 | 97.4 | 84.0 | |
| +LoRA (r=8) | 100.0 | 86.5 | 92.8 | 74.1 | 95.9 | 89.9 | 96.8 | 79.7 | 89.2 | 52.2 | 87.8 | 81.2 | |
| +LoRA (r=32) | 100.0 | 88.5 | 92.9 | 65.7 | 92.9 | 88.0 | 95.7 | 78.7 | 87.4 | 50.1 | 80.5 | 78.5 | |
| +LoRA (r=128) | 100.0 | 89.5 | 93.7 | 60.1 | 83.6 | 85.4 | 86.1 | 77.1 | 77.4 | 43.5 | 58.7 | 68.6 |
This demonstrates that LoRA can be a more parameter-efficient extension based on our method (raw LoRA underperforms DELTA), and can somehow enhance the generality in MKE; however, at the cost of damaging locality. Additionally, we report the time/memory cost in Appendix L (also listed in the response to reviewer QRCS). We have added theoretical analysis and these results into Appendix O. Thanks.
Q4. - Figure 1 has overlapping arrows and labels that are hard to visualize; Figures 2, 3, and 6 are too small to see.
We apologize for these visualization issues. Due to the upload size limit, the first submission compressed the figures heavily (mainly caused by original Figures 2,3, and 6). Now, the revised PDF reduces the number of per-panel cases to restore legibility and maintain clarity. Additionally, vector-format sources with full-resolution figures are already included in the supplementary materials.
Overall, we sincerely thank you for your thoughtful feedback and the considerable time and effort you have devoted to reviewing our work. We would be more than happy to provide any further illustrations or experimental results that may help address your remaining questions or suggestions.
Best,
Authors of Submission 11844.
References
[1] Fang, J., Bi, Z., Wang, R., Jiang, H., Gao, Y., Wang, K., Zhang, A., Shi, J., Wang, X. and Chua, T.S., 2024. Towards neuron attributions in multi-modal large language models. Advances in Neural Information Processing Systems, 37, pp.122867-122890.
[2] Zhou, Y., Cao, P., Chen, Y., Liu, Q., Sui, D., Chen, X., Liu, K. and Zhao, J., 2025, November. M2Edit: Locate and Edit Multi-Granularity Knowledge in Multimodal Large Language Model. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (pp. 29017-29030).
[3] Chen, Q., Zhang, T., Wang, C., He, X., Wang, D. and Liu, T., 2025, April. Attribution analysis meets model editing: Advancing knowledge correction in vision language models with visedit. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 39, No. 2, pp. 2168-2176).
Answer to W1, W2, and W3
Dear Reviewer HVFd:
Thank you for the exceptionally careful reading of our work! We are grateful for the comments help us improve the works. We sincerely hope that our responses and new experiments could address your concerns!
We thank the reviewer for this insightful feedback. In the revised manuscript, we report the causal analysis in Appendix K. The results in Fig. 15 and Fig. 16 show that the result of DELTA localization considerably overlaps with the gaining of causal tracing. This indicates that our decoupled vision-text localization carries genuine causal information rather than merely correlating activation signals. Additionally, the localization results obtained from this causal tracing procedure (Appendix K) have already been used in the experiments in Tab. 3 that combine different localizers with different editors. Under the same editing method, editing causal-tracing-localized layers consistently underperforms editing DELTA-localized layers. These findings suggest that causal relevance and editing suitability are two distinct criteria: while causal tracing identifies fact-related components, DELTA further identifies the components that are more suitable for effective and localized parameter updates.
Thanks for this valuable review, and we appreciate the opportunity to state this precisely. DELTA contributes a more distinct localization and preservation strategy. Based on the illustration in the general response above, we would like to further clarify this as follows.
Prior works address related but different problems. NAM [1] investigates whether MLLM outputs can be attributed to FFN neurons, using image-generation tasks to validate neuron attribution. M2Edit [2] considers language-information propagation and combines important components and layers, without the same fine-grained analysis of internal FFN neurons. In contrast, DELTA is specifically designed for multimodal knowledge editing: it uses a decoupled vision-text localization criterion to identify edit-relevant FFN neurons, rather than simply transferring an attribution or component-selection strategy to the editing setting.
The potential localization-editability alignment. Importantly, identifying components associated with a fact does not necessarily identify the most suitable locations for knowledge editing. As discussed in our response to W1, our causal-tracing analysis (Appendix K, Fig. 15 and Fig. 16) validates that DELTA's localization captures fact-related information. Simultaneously, the localization-editor recombination experiment in Tab. 3 also shows that using causal-tracing-localized layers under the same editor yields worse performance than using DELTA-localized layers. This demonstrates that causal relevance and editing suitability are distinct issues, and our DELTA targets the latter.
The reply to reviewer omH8 and the general response above also provide some illustrations. We updated claims of the contribution in the revision, and we will keep improving the illustrations. Thanks!
We are glad to make clearer explanation: DELTA only requires the VLLMs (i) consists of a decoder-only block with knowledge stored within the post-attention FFN, and (ii) visual signals enter the residual stream within the decoder via visual-token prefixes or cross-attention. Almost all current VLLMs satisfy both conditions, even for architectures with a separate visual memory accessed via decoder cross-attention (e.g., Flamingo). Our DELTA never modifies any vision-side modules (ViT, Q-Former, or projector), which are treated as black boxes. Its localization and edits target the FFN output projection (Wdown). Simultaneously, our experiments have included three LLM families (OPT in BLIP2, LLaMA/Vicuna in LLaVA and MiniGPT4, Qwen2.5/Qwen3-VL) and two visual-injection paradigms (Q-Former vs. patch-level prefixes).