DS@GT's Multilingual Financial QA System at FinMMEval 2026

The team DS@GT from Georgia Tech has developed a retrieval-augmented pipeline for FinMMEval 2026 Task 1, which is a multilingual benchmark for answering financial exam questions in English, Spanish, Greek, Chinese, and Hindi. Their system employs LangGraph to identify the language of queries and retrieves examples from a 30,209-entry multilingual knowledge base utilizing BGE-M3 embeddings and FAISS indexing. Answers are evaluated using Retrieval-Augmented Direct Scoring (RADS), focusing on next-token log-probabilities of candidate letters instead of generating text. For languages with limited resources, the system integrates per-language and cross-lingual retrieval indices through weighted Reciprocal Rank Fusion. Financial certification exams like CFA, EFPA, and CPA necessitate structured reasoning that standard NLP benchmarks do not address, especially in languages with less developed retrieval systems.