Jingze Zhu

Li Li, Yongliang Wu, Jingze Zhu et al.

The evolution of large models has witnessed the emergence of In-Context Learning (ICL) capabilities. In Natural Language Processing (NLP), numerous studies have demonstrated the effectiveness of ICL. Inspired by the success of Large Language Models (LLMs), researchers have developed Large Multimodal Models (LMMs) with ICL capabilities. However, explorations of demonstration configuration for multimodal ICL remain preliminary. Additionally, the controllability of In-Context Examples (ICEs) provides an efficient and cost-effective means to observe and analyze the inference characteristics of LMMs under varying inputs. This paper conducts a comprehensive external and internal investigation of multimodal in-context learning on the image captioning task. Externally, we explore demonstration configuration strategies through three dimensions: shot number, image retrieval, and caption assignment. We employ multiple metrics to systematically and thoroughly evaluate and summarize key findings. Internally, we analyze typical LMM attention characteristics and develop attention-based metrics to quantify model behaviors. We also conduct auxiliary experiments to explore the feasibility of attention-driven model acceleration and compression. We further compare performance variations between LMMs with identical model design and pretraining strategies and explain the differences from the angles of pre-training data features. Our study reveals both how ICEs configuration strategies impact model performance through external experiments and characteristic typical patterns through internal inspection, providing dual perspectives for understanding multimodal ICL in LMMs. Our method of combining external and internal analysis to investigate large models, along with our newly proposed metrics, can be applied to broader research areas.

Jingze Zhu, Yongliang Wu, Wenbo Zhu et al.

Large language models (LLMs) excel at natural language understanding and generation but remain vulnerable to factual errors, limiting their reliability in knowledge-intensive tasks. While decoding-time strategies provide a promising efficient solution without training, existing methods typically treat token-level and layer-level signals in isolation, overlooking the joint dynamics between them. In this work, we introduce a token-aware, layer-localized contrastive decoding method that aligns specific token types with their most influential transformer layers to improve factual generation. Through empirical attention analysis, we identify two key patterns: punctuation tokens receive dominant attention in early layers, while conceptual tokens govern semantic reasoning in intermediate layers. By selectively suppressing attention to these token types at their respective depths, we achieve the induction of controlled factual degradation and derive contrastive signals to guide the final factual decoding. Our method requires no additional training or model modification, and experiments demonstrate that our method consistently improves factuality across multiple LLMs and various benchmarks.