AIAI Seminar – 31 March 2025 – Mengyu Wang and Sunnie Li

Theme: LLMs + Finance

 

Speaker: Mengyu Wang

Title: One More Question is Enough, Expert Question Decomposition (EQD) Model for Domain Quantitative Reasoning

Abstract: Domain Quantitative Reasoning remains a significant challenge for Large Language Models, particularly in fields requiring expert knowledge and complex question answering (QA). This study presents a novel Expert Question Decomposition (EQD) approach to enhance domain-specific reasoning performance. Our method includes two training stages: domain fine-tuning and QA expert alignment, designed to generate targeted supporting questions that optimize the reasoning process. We evaluate our method in the financial domain, an area characterized by specialized knowledge and complex quantitative reasoning requirements. Experimental results across four financial datasets demonstrates the method's effectiveness and efficiency. Our EQD method requires minimal fine-tuning resources, reduces inference time compared to existing prompting methods, and achieves consistent performance improvements ranging from 0.4% to 6.8% across different LLMs. Notably, our analysis reveals that: for LLMs, a single well-crafted supporting question proves more effective than multiple reasoning steps, offering new insights into LLM reasoning mechanisms in specialized domains.

 

Speaker: Sunnie Li

Title: Hierarchical LoRAG, Financial Mapping-Guided Enhanced Answer Retrieval

Abstract: Analyzing financial reports such as 10-K filings presents challenges due to their length, complexity, and domain-specific language, making efficient retrieval difficult for financial NLP systems and retrieval-augmented generation (RAG) models. To address this, we introduce FinGEAR (Financial Mapping-Guided Enhanced Answer Retrieval), a hierarchical retrieval framework for full-document financial analysis. FinGEAR integrates pre-built financial keyword mapping and hierarchical indexing, enabling hybrid navigation that enhances precision and scalability. By leveraging structured document organization and semantic clustering, it efficiently directs queries to relevant sections, minimizing unnecessary retrieval while preserving contextual relevance. Empirical evaluations show that FinGEAR outperforms state-of-the-art retrieval models, improving accuracy and factual consistency to support informed decision-making for analysts, investors, and stakeholders.