Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

MoAI: Mixture of All Intelligence for Large Language and Vision Models

About

The rise of large language models (LLMs) and instruction tuning has led to the current trend of instruction-tuned large language and vision models (LLVMs). This trend involves either meticulously curating numerous instruction tuning datasets tailored to specific objectives or enlarging LLVMs to manage vast amounts of vision language (VL) data. However, current LLVMs have disregarded the detailed and comprehensive real-world scene understanding available from specialized computer vision (CV) models in visual perception tasks such as segmentation, detection, scene graph generation (SGG), and optical character recognition (OCR). Instead, the existing LLVMs rely mainly on the large capacity and emergent capabilities of their LLM backbones. Therefore, we present a new LLVM, Mixture of All Intelligence (MoAI), which leverages auxiliary visual information obtained from the outputs of external segmentation, detection, SGG, and OCR models. MoAI operates through two newly introduced modules: MoAI-Compressor and MoAI-Mixer. After verbalizing the outputs of the external CV models, the MoAI-Compressor aligns and condenses them to efficiently use relevant auxiliary visual information for VL tasks. MoAI-Mixer then blends three types of intelligence (1) visual features, (2) auxiliary features from the external CV models, and (3) language features by utilizing the concept of Mixture of Experts. Through this integration, MoAI significantly outperforms both open-source and closed-source LLVMs in numerous zero-shot VL tasks, particularly those related to real-world scene understanding such as object existence, positions, relations, and OCR without enlarging the model size or curating extra visual instruction tuning datasets.

Byung-Kwan Lee, Beomchan Park, Chae Won Kim, Yong Man Ro• 2024

Related benchmarks

TaskDatasetResultRank
Object Hallucination EvaluationPOPE
Accuracy87.1
2056
Visual Question AnsweringTextVQA
Accuracy67.8
1455
Multimodal UnderstandingMM-Vet
MM-Vet Score43.7
664
Mathematical ReasoningMathVista
Score56.2
566
Multimodal ReasoningMM-Vet
MM-Vet Score43.7
551
Multimodal UnderstandingMMStar
Accuracy48.7
511
Multi-discipline Multimodal UnderstandingMMMU--
422
Science Question AnsweringScienceQA IMG
Accuracy83.5
357
Science Question AnsweringScienceQA SQA-IMG
Accuracy83.5
186
Multimodal Model EvaluationMMBench Chinese
Accuracy76.5
154
Showing 10 of 19 rows

Other info

Code

Follow for update