Ours
Research Journey

"The paper introduces ViInfographicVQA, the first Vietnamese benchmark for Infographic Visual Question Answering, containing 6,747 infographics and 20,409 verified QA pairs across multiple domains. It includes two tasks: Single-image VQA and Multi-image VQA, the latter requiring reasoning across multiple related infographics. Experiments show significant performance gaps in current vision-language models, especially for cross-image reasoning tasks."

"A framework that leverages the region-aware capabilities of the Describe Anything Model (DAM) for Visual Question Answering, particularly in text-rich images. It aggregates answers from multiple regional views of an image to better capture fine-grained textual evidence needed for reasoning. Experiments on six VQA benchmarks show consistent improvements over DAM, including a 7+ point gain on DocVQA, demonstrating strong performance with fewer parameters."

"A Vietnamese version of the VQA-X dataset to support research in Visual Question Answering with Natural Language Explanations. It uses an automated translation pipeline with multiple LLMs to generate and select high-quality Vietnamese translations. Experiments show that NLX-GPT achieves state-of-the-art performance on the new ViVQA-X dataset."

"The paper proposes a Vietnamese VQA training framework using paraphrase-based augmentation and dynamic curriculum learning. Augmented samples are treated as easy and raw samples as hard, with the training ratio gradually adjusted to increase difficulty. Results show improvements on OpenViVQA and mixed results on ViVQA."

"We propose TLVFC, a transfer learning method that enables knowledge transfer between neural networks with different architectures. It initializes convolutional layers via variance-aligned matching of pre-trained weights and fully connected layers using weight distributions, with a feature exchange mechanism during fine-tuning."