Bridging Finance and AI: A Comprehensive Survey of Large Language Models in Financial Applications

Large Language Models (LLMs) have rapidly transformed natural language processing across domains, and their adoption in finance promises to revolutionize tasks ranging from document understanding to quantitative analysis. This survey presents a comprehensive overview of LLM applications in financial contexts. We first trace the evolution of language modeling from RNNs to Transformers and summarize key architectures (encoder-only, decoder-only, encoder-decoder) and transfer-learning paradigms (pretraining, fine-tuning, prompt-based, instruction tuning). We then propose a taxonomy of finance-relevant tasks, including linguistic preprocessing (summarization, NER), sentiment analysis, financial reasoning and QA, forecasting/time-series modeling, and agent-based decision support. Next, we review the landscape of datasets and benchmarks highlighting English-dominant biases, gaps in crisis-period coverage, and emerging multilingual resources. We catalog financial domain--specific LLMs, contrasting fine-tuned models (FinBERT, FinGPT, FinMA) with from-scratch systems (BloombergGPT, FinTral, XuanYuan 2.0) and compare them to general-purpose models. We examine evaluation metrics, contrasting standard ML measures (accuracy, F1, MSE) with finance-centric criteria (Sharpe Ratio, Maximum Drawdown) and discuss challenges around data leakage, hallucinations, robustness, and explainability. Ethical and regulatory considerations data privacy, bias, fairness, and compliance are addressed, followed by a multi-level decision framework (zero-shot through pretraining from scratch) that balances accuracy, cost, and privacy. Finally, we outline open challenges and future directions, including cross-lingual models, symbolic reasoning integration, human--AI collaboration, and the development of open, temporally annotated finance benchmarks. This survey aims to guide researchers and practitioners in responsibly harnessing LLMs to advance financial AI.