GPT-4.1 Mini tops financial causality QA with multilingual fine-tuning

This paper from Team HSA_CORAL tackles the FinCausal 2026 shared task, which focuses on extracting cause-effect relations from financial narratives using extractive question answering in both English and Spanish. The team systematically compared three model families: encoder-only token tagging with multilingual BERT, encoder-decoder generation with multilingual BART, and decoder-only LLMs including Llama 3.1 and GPT variants (e.g., GPT-4.1 Mini). They explored prompt refinement, few-shot demonstrations, and supervised fine-tuning across all models. The key finding was that while prompting and few-shot examples delivered competitive baselines, supervised fine-tuning provided the largest performance gains, especially when combining data from both languages.

The top-performing system was GPT-4.1 Mini fine-tuned on the merged English and Spanish training data. Under the shared task’s novel LLM-as-a-judge evaluation metric, this model achieved a tied highest score of 4.8140 on the English subtask and ranked third on Spanish with a score of 4.7753. The results demonstrate that task-specific adaptation, particularly through multilingual fine-tuning, dramatically improves cross-lingual transfer for financial causality QA. This work underscores the practical value of fine-tuning smaller, efficient LLMs on domain-specific multilingual corpora, achieving state-of-the-art performance without resorting to massive model sizes or extensive prompt engineering.