Dynamic Adapter Routing for Continual Multimodal Retrieval
A recent study presents Dynamic Adapter Routing (DAR), a technique designed for ongoing multimodal retrieval in vision-language models. The researchers contend that current methods, typically reliant on class-incremental learning (CIL), do not perform well in practical scenarios. They introduce an innovative evaluation framework for continual multimodal retrieval (CMR) that encompasses various visual domains. By employing prototype-based routing to select adapters and merging models, DAR demonstrates enhanced performance compared to earlier benchmarks. This research can be accessed on arXiv with the identifier 2605.31229.
Key facts
- Dynamic Adapter Routing (DAR) is proposed for continual multimodal retrieval.
- Existing CIL methods fail in the proposed more challenging scenario.
- A new evaluation framework for CMR is introduced.
- DAR uses prototype-based routing and model merging.
- The paper is on arXiv with ID 2605.31229.
- The approach targets vision-language models.
- Continual retrieval remains underexplored.
- DAR achieves superior performance over baselines.
Entities
Institutions
- arXiv