PhD Seminar • Information Retrieval | Machine Learning • UniRAG: Universal Retrieval Augmentation for Multi-Modal Large Language Models

Please note: This PhD seminar will take place in DC 1304.

Sahel Sharifymoghaddam, PhD candidate
David R. Cheriton School of Computer Science

Supervisor: Professor Jimmy Lin

Recently, Multi-Modal (MM) Large Language Models (LLMs) have unlocked many complex use-cases that require MM understanding (e.g., image captioning or visual question answering) and MM generation (e.g., text-guided image generation or editing) capabilities. To further improve the output fidelity of MM-LLMs we introduce the model-agnostic UniRAG technique that adds relevant retrieved information to prompts as few-shot examples during inference. Unlike the common belief that Retrieval Augmentation (RA) mainly improves generation or understanding of uncommon entities, our evaluation results on the MSCOCO dataset with common entities show that both proprietary models like GPT4 and Gemini-Pro and smaller open-source models like Llava, LaVIT, and Emu2 significantly enhance their generation quality when their input prompts are augmented with relevant information retrieved by MM retrievers like UniIR models.