Vision-Language Models (VLMs) have demonstrated remarkable proficiency in general multi-modal understanding; yet they struggle to efficiently acquire continually evolving domain-specific skills. Conventional approaches to enhancing VLM capabilities, such as Supervised Fine-Tuning (SFT), require extensive dataset curation and substantial computational resources. Model merging has emerged as an efficient alternative that enables the transfer of domain-specific expertise from Large Language Models (LLMs) to VLMs without incurring additional training data requirements or significant computational overhead. Unlike conventional merging of homogeneous LLMs, which mainly aggregates existing capabilities, cross-modal skill injection aims to induce emergent cross-modal capabilities by integrating a domain-expert LLM into a VLM. However, existing research lacks a systematic analysis of the applicability and methodology of cross-modal skill injection. In this study, we investigate cross-modal skill injection across three main aspects: scenarios, methods, and hyperparameters. For scenarios, we find that cross-modal skill injection generally performs well in instruction-following and cross-lingual settings, yet struggles with mathematical reasoning. For methods, we find that classic approaches such as TA and DARE consistently achieve superior performance over alternative merging methods. We also provide a systematic and quantitative analysis of the hyperparameter tuning that these classic methods critically depend on.

Accurate trading volume prediction is essential for portfolio optimization, market regulation, and financial risk control. An effective method for predicting trading volume involves building a graph to model relations between stock. Recent research has enhanced these models by integrating stock news to improve forecasting ability. However, existing approaches primarily integrate news data as auxiliary features for nodes in Graph Neural Networks (GNNs), overlooking the relational information between stocks embedded in news. To address this, we propose LLM-Enhanced Dynamic Graph Neural Network (LED-GNN), a framework that constructs dynamic graphs using inter-stock relationships extracted from news via a large language model (LLM)-centered pipeline, combined with graphs learned from historical price-volume data. A dynamic GNN then processes these graphs to generate predictions. Evaluated on a real-world dataset, TOPIX, with Reuters Financial News, LED-GNN consistently outperformed all baseline models, achieving a 2% improvement over the strongest baseline.