ARMDIL: MLLM-Routed Heterogeneous Ensembles for Robust Cross-Dataset Image Classification
A new study on arXiv (2608.13463) introduces ARMDIL, an Adaptive Router aimed at Multi-Domain Image classification using LLMs. This advanced system features a multimodal large language model (MLLM) agent that smartly assigns images to the best-suited vision backbone from a diverse array of convolutional neural networks (like ResNets), self-supervised learners (SSL), and vision-language models (VLMs). These models are built on a unified label space created from different image datasets with distinct features. The findings show that ARMDIL effectively navigates the trade-offs among various architectures, rivaling routers that rely on specialized training. The research highlights the strengths and limitations of each architecture in different visual contexts, addressing the challenge of image classification models that excel in one dataset but struggle in others. By leveraging multiple model types, ARMDIL enhances cross-dataset classification.
Key facts
- ARMDIL is an Adaptive Router for Multi-Domain Image classification with LLMs.
- It uses a multimodal large language model (MLLM) agent to route images to suitable vision backbones.
- The ensemble includes ResNets, self-supervised representation learners (SSL), and vision-language models (VLMs).
- Models are trained on a unified label space from multiple image datasets.
- ARMDIL performs competitively with specialized training-based routers.
- The paper is available on arXiv with ID 2608.13463.
- The research addresses generalization across domains and difficulty levels.
- Empirical evaluations show distinct capabilities and vulnerabilities of each architecture.
Entities
Institutions
- arXiv