• Authors: Yichi Zhang, Zhuo Chen, Lingbing Guo, Lei Liang, Wen Zhang, Huajun Chen

Understanding and reasoning with abstractive information from the visual modality presents significant challenges for current multi-modal large language models (MLLMs). Among the various forms of abstractive information, Multi-Modal Relational Knowledge (MMRK), which represents abstract relational structures between multi-modal entities using node-edge formats, remains largely under-explored. In particular, STructured and Abstractive Reasoning (STAR) on such data has received little attention from the research community. To bridge the dual gaps in large-scale high-quality data and capability enhancement methodologies, this paper makes the following key contributions: (i). An automatic STAR data engine capable of synthesizing images with MMRK to build multi-modal instruction data with reliable chain-of-thought thinking for various STAR tasks and (ii). A comprehsive two-stage capability enhancement training framework, accompanied by a suite of evaluation protocols tailored to different STAR tasks. Based upon these contributions, we introduce STAR-64K, a dataset comprising 64K high-quality multi-modal instruction samples, and conduct experiments across 5 open-source MLLMs. Experimental results show that our two-stage enhancement framework enables smaller 3B/7B models to significantly outperform GPT-4o in STAR. Additionally, we provide in-depth analysis regarding the effectiveness of various designs, data transferability, and scalability.

• The full paper, data, and code would be released in the future.

Many images contain abstractive high-level semantic information that is artificially defined and does not exist in nature. Teaching MLLMs to understand and reason about this abstractive information is a significant challenge. In this work, we introduce a novel type of abstractive image data: multi-modal relational knowledge images.

Here is an overview of our data engine, the training pipeline, the seed tasks, and the CoT prompts.

• First, you should install LLaMA-Factory and vLLM in your python environment.
• Second, you need to download the MLLMs used in the experiments including Qwen2.5-VL-3B/7B/32B, LLaVA-1.5-7B, and LLaVA-NEXT-34B
• Next, run bash train.sh to fine-tune MLLMs with LLaMA-Factory.
• Finally, use vLLM to conduct inference on the trained MLLMs to obtain the results and calculate the metrics.
• The full code would be released in the future.