Language-Routed RAG and Direct Option Scoring for Multilingual Financial QA: DS@GT at FinMMEval

This paper describes DS@GT’s submission to FinMMEval 2026 Task 1, a multilingual financial exam QA benchmark covering English, Spanish, Greek, Chinese, and Hindi. The system detects the query language, retrieves exemplars from a 30,209-entry multilingual knowledge base using BGE-M3 and FAISS, and applies Retrieval-Augmented Direct Scoring (RADS) by comparing next-token log-probabilities for answer-option letters. Models are routed by language: Qwen3-14B for Arabic, Chinese, and Hindi; Qwen2.5-14B for English; and Llama-3.1-8B for Greek. Reported ablations expose severe language- and decoding-strategy asymmetries.