- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies lifelong multimodal knowledge editing and proposes MoAE, an extension of MoE-based editing (ELDER) with dynamic expert expansion. Specifically, (1) a new expert is added when the distance between an input representation and existing routers exceeds a threshold ; (2) each expert consists of a shared LoRA and multiple specific LoRAs, where a new specific LoRA is added every samples processed by that expert; and (3) a replay-based margin loss is used to prevent router drift, combined with an external image-text retrieval step (DINOv2 + reranker) for inference gating. The method is evaluated on ComprehendEdit and VLKEB.
Critical strengths
- The motivation is practical. Since pre-defining a fixed number of experts and LoRA ranks is difficult for streaming edits of unknown length (Fig. 1a), expanding experts by representation distance and adding specific LoRAs by usage count is a sensible design.
- The paper identifies the "router forgetting" issue when adding new routers online and addresses it by replaying boundary representations with a margin loss (Sec. 3.2.1), which is shown to be effective in Table 3.
Critical weaknesses
- Unfair inference retrieval setup inflates Locality and masks Generality drops (Sec. 3.2.3). The memory bank stores pre-extracted DINOv2 and text-reranker features for all edited samples along with of unedited locality samples from the training set. At test time, if the nearest neighbor in is a locality sample, the model skips the edited modules entirely and returns the original model . This trivially gives 100.00 T-L in Table 1 and makes the baseline comparison unfair. Meanwhile, on Text Generality (T-G), where the edited modules are actually used, MoAE drops noticeably behind ELDER (e.g., ComprehendEdit : 66.00 vs. 78.26; VLKEB : 28.42 vs. 46.25). Its higher AVG score (Eq. 4) is mainly driven by averaging in the saturated locality metrics.
- Mismatch between and across layers, plus copy-paste errors in Table 1. Sec. 3.2 triggers expansion at layer using layer- features and routers , but Sec. 3.2.1 and Eq. (3) state that only stores a single representation per sample to compute router weights across all injected layers (22–26). Since feature spaces change across layers, it is unclear how can be directly compared with deeper routers during routing and replay. Also, in Table 1, all six metrics for
Merging() andSD-LoRA() are identical across ComprehendEdit and VLKEB, which looks like a copy-and-paste error.
Questions for authors
-
What are the T-L, M-L, and KPI scores of MoAE if you remove the external DINOv2 + reranker retrieval step and the 20% stored locality samples in Sec. 3.2.3, and use only the original deferral mechanism from ELDER?
-
Could you clarify the discrepancy between and ? In which feature space do the routers (for layers 22–26) reside, and how is router replay (Eq. 2) computed for deeper layers if only is stored?
-
Why are the results for
Merging() andSD-LoRA() in Table 1 identical across ComprehendEdit and VLKEB?