Reproduce “Large Vision-Language Model Alignment and Misalignment: A Survey Through the Lens of Explainability” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dong Shu, Haiyan Zhao, Jingyu Hu, Weiru Liu, Ali Payani, Lu Cheng, Mengnan Du
This survey paper establishes a comprehensive framework for understanding alignment and misalignment in Large Vision-Language Models (LVLMs), defining alignment across representational and behavioral aspects. It proposes a detailed taxonomy of misalignment (object, attribute, relational), analyzes its root causes at the dataset, model, and inference levels, and systematically categorizes mitigation strategies through an explainability lens.
Key insights
- LVLM alignment is a dual concept, encompassing both the semantic correspondence of internal representations and the fidelity of generated outputs, facilitated by multi-stage training.
- Misalignment extends beyond simple 'hallucination' and can be precisely categorized by the type of semantic divergence: incorrect objects, attributes, or relationships in model descriptions.
- Root causes of misalignment are multifaceted, stemming from issues in training data quality, inherent architectural limitations or knowledge conflicts within the model, and challenges during the inference process, such as out-of-distribution tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.