Synthetic Contrastive Reasoning for Multi-Table Q&A
Researchers have developed a new method for multi-table question answering by creating a synthetic dataset of reasoning traces. This dataset, generated using large language models, includes both correct and plausible incorrect reasoning paths. Fine-tuning open-weight models like Qwen3-14B, Mistral-8B, and Llama-3.1-8B with this contrastive data significantly improved their question-answering performance compared to standard supervised fine-tuning. AI
IMPACT Introduces a novel dataset and fine-tuning technique to improve LLM performance on complex relational data reasoning tasks.