Introduction
As Large Language Models (LLMs) transition from general-purpose assistants to specialized tools, they encounter the unique "high-stakes" requirements of Finance, Healthcare, and Law (FHL). In these domains, the cost of a "hallucination" is not merely a conversational quirk but can lead to significant financial loss, medical malpractice, or legal injustice. This survey maps the current landscape of "LLM+X" in these sectors, examining how domain-specific adaptations address professional expertise, data privacy, and regulatory compliance.
The FHL domains share a unique set of constraints that distinguish them from general NLP applications:
- Professional Expertise: Tasks often require specialized training, such as interpreting complex tax codes or diagnostic imaging.
- Confidentiality: Data is frequently protected by regulations like HIPAA or KYC, limiting the use of open-source training data.
- Multimodality: Information is often embedded in non-textual formats, including tables, charts, and medical scans.
- Strict Regulation: Outputs must be explainable, fair, and legally compliant.
Methodological Framework
The evolution of AI in FHL domains can be categorized into three primary methodological waves:
- Domain-Specific Pre-training: Models like BERT or RoBERTa are trained from scratch or continued on massive domain-specific corpora (e.g., BloombergGPT for finance).
- Instruction Fine-Tuning (IFT): General LLMs are adapted using specialized instruction sets to follow domain-specific logic (e.g., FinMA or ChatLaw).
- Advanced Prompting & RAG: Techniques such as Chain-of-Thought (CoT), Program-of-Thought (PoT)—where models generate code for numerical tasks—and Retrieval-Augmented Generation (RAG) provide factual context without retraining.
Domain-Specific Benchmarks and Tasks
The survey identifies a wide array of benchmarks across the three sectors. In Finance, tasks range from sentiment analysis (FPB) to numerical reasoning over tables (FinQA). Healthcare benchmarks include MedQA (USMLE-style) and MIMIC-CXR for radiology. Legal benchmarks like LegalBench and CAIL2018 focus on reasoning and judgment prediction.
Key Insights and Sector Performance
A central argument of this survey is the "FHL Synergy": these domains should be studied together due to their shared "high-stakes" profile, necessitating a shift from accuracy-only metrics to comprehensive "trustworthiness" evaluations.
- Finance: Numerical reasoning remains a critical frontier. Models increasingly utilize "Program-of-Thought" prompting to ensure calculation accuracy. While GPT-4 leads in complex reasoning, instruction-tuned models like FinMA-30B are narrowing the gap.
- Healthcare: Specialized models like Med-PaLM 2 have achieved expert-level performance on medical exams. Multimodal Vision-Language Models (VLMs) are also emerging to assist in clinical report generation.
- Law: Despite passing the Bar Exam, LLMs still struggle with "Legal Reading Comprehension" in long-form documents compared to specialized supervised baselines.
Challenges and Future Directions
Several bottlenecks remain for the next generation of FHL AI:
- Evaluation Gap: Current multiple-choice benchmarks do not reflect real-world professional workflows; more free-form evaluation frameworks are needed.
- Hallucination Mitigation: In legal and medical contexts, hallucinations remain a critical safety risk that RAG only partially addresses.
- Human-AI Collaboration: Research is shifting toward "Human-in-the-loop" models where AI assists rather than replaces experts.
- Data Scarcity: Privacy constraints necessitate the exploration of synthetic data generation and privacy-preserving techniques like Federated Learning.
Summary updated