Reproduce “NoiseBoost: Alleviating Hallucination with Noise Perturbation for Multimodal Large Language Models” 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

Kai Wu, Boyuan Jiang, Zhengkai Jiang, Qingdong He, Donghao Luo, Shengzhi Wang, Qingwen Liu, Chengjie Wang

NoiseBoost introduces a training method for Multimodal Large Language Models (MLLMs) that perturbs visual features with noise to mitigate hallucination. This approach improves factual accuracy and the balanced distribution of attention, leading to more reliable long-form content generation and enabling the first reported semi-supervised learning for MLLMs without increasing inference time or data labeling costs.

Key insights

  • MLLM hallucinations originate from the LLM's inherent summarization mechanism and over-reliance on linguistic 'anchor tokens,' causing neglect of comprehensive visual context.
  • Perturbing visual features during training compels MLLMs to distribute attention more evenly between visual and linguistic tokens, reducing overconfidence in specific linguistic patterns and fostering robust visual understanding.
  • The benefits of perturbation are specific to visual tokens; pixel-level or color-space distortions generally degrade performance, and perturbing language embeddings negatively impacts the strong pre-trained knowledge in foundation LLMs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.