ITUNLP at IberLEF-PRESTA: A Zero-Shot Code Generation Approach for Question Answering over Spanish Tabular Data

This paper presents our zero-shot, LLM-driven code generation approach for solving the IberLEF 2025 - PRESTA: Question Answering over Tabular Data in Spanish task. Our approach relies on a Python code generation framework that employs state-of-the-art large language models (LLMs), including OpenAI o3, Qwen3, DeepSeekR1, DeepSeek-V3, Llama 4, to generate executable Pandas code via optimized prompting strategies. Experimental results show that different LLMs vary in their effectiveness for code generation, and our hybrid configuration achieved the highest accuracy among the seven participating teams in the shared task. Specifically, our system reached 90% accuracy on the development set and 87% on the test set, demonstrating the viability of zero-shot methods for tabular question answering.