acceptodds
Withdrawn from ICLR 2027

EdgeEdit: An Edge-Oriented Multimodal Knowledge Editing Framework for On-Device MLLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

On-device multimodal large language models (MLLMs) are increasingly enabling multimodal applications, yet updating their static encoded knowledge efficiently remains a critical challenge. In-context editing (ICE) provides an efficient and lightweight approach for fine-grained knowledge updates. However, on-device deployment typically demands joint model and input compression, introducing severe obstacles for ICE: model quantization impairs the model’s ability to utilize contextualized edited knowledge, while edit-agnostic token reduction inadvertently discards edit-critical visual evidence. To address these challenges, we propose EdgeEdit, an edge-oriented multimodal knowledge editing framework that coordinates offline model deployment with online inference. Specifically, EdgeEdit introduces Edit-Aware Mixed-Precision Quantization for offline model deployment, which conducts zeroth-order sensitivity profiling and budget-constrained dynamic programming to preserve contextualized edited knowledge utilization. For online inference, EdgeEdit employs EvidenceSaliency Guided Token Reduction, combining cross-modal saliency attribution with semantic relevance to progressively remove redundant visual tokens while retaining edit-critical visual evidence. Extensive experiments demonstrate that EdgeEdit maintains competitive editing performance while substantially reducing system costs. Compared with the strongest competing baseline for each efficiency metric, EdgeEdit achieves average reduction factors of 2.9×/1.5×/2.5× in memory usage, editing latency, and energy consumption on the server, and 2.0×/2.2×/2.0× on mobile devices, respectively.

Primary area:
foundation or frontier models, including LLMs
Keywords:
On-device Multimodal Large Language Models, Multimodal Knowledge Editing, In-Context Editing, Mixed-Precision Quantization, Visual Token Reduction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_96gSOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies multimodal knowledge editing for resource-constrained on-device MLLMs and proposes EdgeEdit, an edge-oriented framework that jointly considers model compression and input compression for efficient in-context editing. Specifically, the paper identifies two compression-related challenges: quantization can degrade the utilization of retrieved edited knowledge, while edit-agnostic visual token reduction may remove edit-critical visual evidence. To address these issues, EdgeEdit introduces an edit-aware mixed-precision quantization strategy and an evidence-saliency guided token reduction module. The paper evaluates EdgeEdit on multimodal knowledge editing benchmarks with both server-side and mobile-device settings, providing a relatively novel experimental perspective by considering editing effectiveness together with practical system metrics such as memory usage, latency, and energy consumption. However, the technical contributions are mainly based on adapting existing compression techniques, including sensitivity-aware mixed-precision quantization and visual token reduction, to the knowledge editing scenario. Therefore, while the problem setting and evaluation perspective are interesting and practically meaningful, the methodological novelty is relatively limited and the work is more characterized as an engineering-oriented optimization framework rather than a fundamentally new algorithmic contribution.

Critical strengths
  1. The work extends multimodal knowledge editing from conventional server-side evaluation to resource-constrained on-device MLLM deployment, where practical constraints such as memory, latency, and energy consumption are considered together with editing quality. This provides a valuable perspective for real-world deployment and makes the evaluation setting more realistic than existing MKE studies.
  2. The paper evaluates the proposed framework across multiple MLLM backbones, editing benchmarks, and hardware platforms, while reporting both editing metrics and system-level efficiency metrics. Although the algorithmic novelty is limited, the extensive system evaluation demonstrates the practical effectiveness of applying compression-aware optimization to multimodal knowledge editing and supports the empirical claims of the paper.
Critical weaknesses
  1. The technical novelty of EdgeEdit is somewhat incremental. EAMPQ follows the sensitivity-aware mixed-precision quantization paradigm, similar to AWQ and KVTuner [1,2], while ESGTR is closely related to existing visual token reduction methods such as SparseVLM [3]. The main difference is introducing editing-aware objectives into these existing compression frameworks. Therefore, the contribution appears to be primarily an adaptation of existing quantization and token reduction techniques to the knowledge editing scenario rather than a fundamentally new compression methodology.
  2. The practical effectiveness for real-world on-device editing remains insufficiently validated. Although EdgeEdit targets resource-constrained deployment, the paper does not fully characterize the practical deployment cost and scalability. In particular, the offline overhead introduced by EAMPQ profiling is not reported, and the mobile efficiency comparison may be affected by the use of NPU acceleration for EdgeEdit while baselines run on CPU. Moreover, the evaluation mainly focuses on individual editing requests and does not investigate realistic scenarios with sequential edits or growing retrieval contexts, which are important for long-term on-device knowledge editing.

[1] LIN J, TANG J, TANG H, et al. AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration[C]//Proceedings of Machine Learning and Systems. 2024.

[2] LI X, XING Z, LI Y, et al. KVTuner: Sensitivity-aware Layer-wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference[C]//International Conference on Machine Learning. 2025.

[3] ZHANG Y, FAN C K, MA J, et al. SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference[C]//International Conference on Machine Learning. 2025.

Questions for authors

1.Could the authors clarify the key novelty of EdgeEdit compared with existing compression-based optimization methods? 2.Could the authors provide the offline profiling cost of EAMPQ? 3.Could the authors clarify the contribution of NPU acceleration to the reported mobile efficiency gains? 4.Could the authors provide results under sequential editing settings?

Reviewer_UXxnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multimodal in-context knowledge editing on resource-constrained devices. It builds on HICE and introduces two compression modules to allocate the precision budget across layers and reduce visual tokens. EAMPQ assigns quantization precision to Transformer layers using calibration losses on editing and unrelated queries. ESGTR progressively reduces visual tokens using attention, semantic similarity, and token merging. Experiments on two models and two editing datasets report better editing performance and lower memory use, latency, and energy consumption on servers and mobile phones.

Critical strengths
  • The problem has practical value. Enabling multimodal models on phones to use new knowledge without retraining the main model for each update is worth studying.

  • Using editing tasks to guide quantization calibration is reasonable. Table 4 shows that, under the same mixed-precision budget, EAMPQ improves reliability and textual generality over generic precision allocation by 6.18 and 3.46 percentage points, respectively.

  • The overall pipeline is clear. The experiments use two models and two datasets, with measurements on both servers and phones. Table 3 also shows that the two modules together can maintain or improve editing quality under compression.

Critical weaknesses
  • The mobile comparisons use different execution backends. In Figure 5, EdgeEdit runs on the NPU, while the baselines run on the CPU. Compression, the execution backend, and the amount of disk offloading all affect latency and energy consumption. The comparison therefore cannot show how much of the gain comes from compression and how much comes from hardware acceleration. The methods should be compared on the same backend under the same resource budget.

  • The motivation for editing-specific compression is not clear. In Table 6, GPTQ-W4A16 already performs close to the uncompressed model on E-VQA. In Table 5, EATR itself improves performance over no compression. ESGTR adds only another 0.58 and 0.44 percentage points in reliability and textual generality, respectively. These results show that good general-purpose compression methods are already sufficient for this type of task.

  • The paper does not clearly show what the new design choices add. EAMPQ's main addition is its calibration objective for editing, while ESGTR combines several token scoring and merging strategies. Table 5 only compares the scoring rule against attention from the last prompt token. Without comparisons to related methods such as SparseVLM under the same budget, it is hard to judge the value of the new design choices.

  • The calibration and deployment details are incomplete. The paper does not clearly state the calibration set size or the split between calibration and evaluation data. It is therefore unclear whether the precision configuration works on unseen edits. The actual quantization configuration, token reduction settings, NPU execution, and energy measurement procedure are also not described in enough detail. The offline calibration cost is not reported. These omissions make it difficult to reproduce the results and assess the cost.

Questions for authors
  • Can the authors compare EdgeEdit with HICE using standard quantization and token reduction on the same mobile backend, under comparable resource budgets?

  • Why do EAMPQ and ESGTR improve editing quality over uncompressed HICE? For example, in Table 3, reliability and textual generality on ComprehendEdit increase from 78.09/69.81 to 82.95/75.12, and each module also improves performance when used alone. Please explain why quantization and token reduction each lead to this improvement.

  • How is the calibration set constructed? Are its editing instances, generality queries, and locality queries separate from the evaluation data? Is the precision configuration fixed before evaluating unseen edits?

  • Is locality measured against the uncompressed full-precision model, or against each compressed model before editing? Do unrelated queries go through the same compressed inference pipeline? This matters for interpreting the nearly unchanged locality scores in Tables 3 and 6.

Reviewer_tRCKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents EdgeEdit, a framework for efficient multimodal in-context knowledge editing (ICE) on resource-constrained devices. The authors identify two challenges introduced by model and input compression: quantization can impair the utilization of contextualized edited knowledge, while edit-agnostic token reduction may discard relevant visual evidence. EdgeEdit addresses these issues through two components: Edit-Aware Mixed-Precision Quantization (EAMPQ), which uses zeroth-order sensitivity profiling and dynamic programming for layer-wise precision allocation, and Evidence-Saliency Guided Token Reduction (ESGTR), which combines cross-modal attention and semantic relevance for progressive visual-token reduction. Experiments on E-VQA and ComprehendEdit with two MLLM backbones show competitive editing performance alongside reduced memory consumption, latency, and energy usage.

Critical strengths
  1. Well-motivated problem. The paper addresses the practical challenge of efficient multimodal knowledge editing on resource-constrained devices. The analysis of quantization-induced degradation and visual evidence loss provides clear motivation for the proposed approach.

  2. Coherent framework design. EdgeEdit combines edit-aware mixed-precision quantization with evidence-guided token reduction. The two components target distinct compression challenges and form a complementary offline–online framework.

  3. Solid experimental evaluation. The experiments cover two benchmarks, two MLLM backbones, and three mobile devices. The ablation results support the effectiveness of both components, particularly the gains of EAMPQ over generic mixed-precision quantization.

Critical weaknesses
  1. Unfair mobile efficiency comparison. EdgeEdit uses NPU acceleration, while competing methods run on CPUs. This makes it difficult to isolate the efficiency gains from the proposed algorithms rather than hardware differences.

  2. Limited evaluation of unseen knowledge edits. EAMPQ relies on edit-aware calibration data, but the paper does not clearly establish how well the resulting precision allocation generalizes to new edits after deployment. The offline profiling overhead is also insufficiently evaluated.

  3. Modest gains from token reduction. ESGTR only slightly outperforms edit-agnostic token reduction in Table 5. More detailed component ablations and variability analysis are needed to better establish the effectiveness of its saliency design.

  4. Incomplete coverage of related work. The paper overlooks several relevant multimodal knowledge editing studies, including M-ORE, ScopeEdit, M2Edit, and MC-MKE. A more comprehensive discussion would help better position EdgeEdit within recent developments in online editing, cross-modal generalization, and modality consistency.

Questions for authors
  1. Can the authors provide a fairer mobile comparison using the same execution backend, or quantify how much of the efficiency gain comes from NPU acceleration rather than the proposed methods?

  2. How well does EAMPQ generalize to unseen knowledge edits after deployment? Is the calibration data disjoint from the evaluation set, and what is the associated offline profiling cost?

  3. The improvements of ESGTR over EATR in Table 5 are relatively small. Are these gains consistent across different settings, and which saliency components contribute most to the improvement?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.