DS@GT's Language-Routed RAG Boosts Multilingual Financial QA Accuracy

Researchers from Georgia Tech (DS@GT) have developed a multilingual financial question-answering pipeline for the FinMMEval 2026 benchmark, covering English, Spanish, Greek, Chinese, and Hindi. The system, called Language-Routed RAG with Direct Option Scoring, addresses the challenge of structured domain reasoning in financial certification exams (CFA, EFPA, CPA) across languages with underdeveloped retrieval infrastructure. Their pipeline (built on LangGraph) detects query language and retrieves semantically relevant exemplars from a 30,209-entry multilingual knowledge base using BGE-M3 embeddings and FAISS indexing. Instead of generating free-form answers, the system uses Retrieval-Augmented Direct Scoring (RADS) by reading next-token log-probabilities over candidate option letters. For low-resource languages, the team fuses per-language and cross-lingual retrieval indices via weighted Reciprocal Rank Fusion.

The most striking findings involve the fragility of reasoning methods across languages. Model routing is language-dependent: Qwen3-14B for Arabic, Chinese, and Hindi; Qwen2.5-14B for English; and Llama-3.1-8B for Greek. Ablation experiments revealed severe language-asymmetric performance gaps. Chain-of-thought prompting, often considered beneficial, dramatically degraded Greek accuracy from 90.7% to 20.9%. Similarly, enabling Qwen3's default thinking mode collapsed Arabic RADS performance to near-chance levels. These results indicate that effective multilingual financial reasoning requires language-aware retrieval, careful model routing, and deliberate scoring strategy selection. The paper has been accepted for publication in the CLEF 2026 Working Notes.