Multimodal Visual RAG for Reducing MLLM Hallucinations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multimodal Large Language Models (MLLMs) are prone to visual hallucinations and visual distractions, particularly in healthcare scenarios, leading to inaccurate medical diagnoses and treatment plans due to limited domain knowledge and reliance on text-based retrieval-augmented generation techniques.
Innovation Solution
Fine-tuning MLLMs with awareness, focus, and learning datasets to enhance image-text association, minimize distractions, and mitigate hallucinations by incorporating both visual and textual modalities in retrieval-augmented generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If text-based retrieval-augmented generation techniques are used, then the model can generate responses, but visual hallucinations and inaccuracies increase due to limited domain knowledge
Solution Approach 1:
The patent merges text-based retrieval-augmented generation with vision-based processing by combining textual queries with retrieved images. The system integrates both modalities to enhance the model's understanding and reduce hallucinations, as the visual information provides additional grounding for the generated responses.
Solution Approach 2:
The patent introduces an intermediary mechanism that retrieves relevant images based on textual queries and uses them to augment the generation process. This intermediary step of image retrieval and integration helps the model ground its responses in visual evidence, reducing hallucinations while maintaining accuracy.
2Loss of information
If randomly chosen images are added to each example from the relevant dataset, then the model learns image-text associations, but visual distractions increase
Solution Approach 1:
The patent applies local quality by distinguishing between different types of images in the dataset. Instead of treating all images equally, it identifies and prioritizes relevant images that match the query while filtering out distracting images. This selective approach maintains image-text association capability while minimizing visual distractions.
Solution Approach 2:
The patent segments the dataset into different categories or subsets based on relevance to the query. By dividing the images into relevant and irrelevant segments, the model can learn image-text associations from relevant images while excluding distracting images from the learning process, thus reducing visual distractions.
3Measurement precision
If finetuning is performed with awareness and focus datasets, then the model processes images more accurately, but processing time increases
Solution Approach 1:
The patent applies preliminary action by pre-processing and organizing the dataset into awareness and focus subsets before the actual processing task. This pre-organization allows the model to efficiently navigate and process only the relevant information during inference, reducing processing time while maintaining accuracy through the structured approach.
Solution Approach 2:
The patent uses partial action by selectively processing only the necessary portions of the dataset (awareness and focus datasets) rather than processing the entire dataset. This selective processing approach maintains image processing accuracy for relevant images while reducing the overall processing time by excluding irrelevant data from the fine-tuning process.
Data Source
AI summary
Systems and methods for visual retrieval augmented generation for artificial intelligence models such as multimodal large language models. Associations between image and description pairs can be identified from an awareness dataset by finetuning a multi-modal large language model (MLLM) with the awareness dataset based on randomly chosen images added to each example from a relevant dataset. Visual distractions for image processing with the MLLM can be minimized by finetuning the MLLM with a focus dataset based on randomly chosen images added to each example from the relevant dataset. Visual hallucinations from the MLLM can be mitigated by finetuning the MLLM with a learning dataset based on related images having corresponding texts added to each example from the relevant dataset to utilize extracted information from associations between provided text from multiple images and a learning dataset.