Robustness of Multimodal RAG Systems

Establish robustness enhancement techniques for multimodal retrieval-augmented generation systems that mitigate modality bias and adversarial perturbations and preserve performance when confronted with low-quality or outdated external sources.

Background

The paper notes that multimodal RAG systems often over-rely on text, struggle with precise source attribution, and are vulnerable to adversarial inputs and degraded sources.

While unimodal RAG trustworthiness has been studied, the robustness of multimodal RAG remains unresolved and is identified as a key research priority.

References

While the trustworthiness of unimodal RAGs has been studied , enhancing the robustness of multimodal RAGs remains an open challenge and a promising research direction.

— Ask in Any Modality: A Comprehensive Survey on Multimodal Retrieval-Augmented Generation  (2502.08826 - Abootorabi et al., 12 Feb 2025) in Section 6, Open Problems and Future Directions — Generalization, Explainability, and Robustness

Self-RAG [13] adds reflection tokens for retrieval and critique, but it is unclear whether self-correction can resist adversarial inputs designed to manipulate the critique process.

— Trustworthy RAG: An Evaluation Agent for Detecting Misinformation and Knowledge Poisoning in Generative AI Systems  (2608.21095 - Giri et al., 21 Aug 2026) in Section II, Background and Related Work, p. 2

Evaluating $imMRAG$ against a defended system, and establishing how far the embedded instruction can be obfuscated while remaining legible, is the most consequential direction left open by this work.

— Walking the Embedding Space: Datastore Extraction from Multimodal RAG  (2610.01871 - Jica et al., 1 Oct 2026) in Section 5, paragraph "Realism of the setting" in Section "Limitations"