{"ID":3084652,"CreatedAt":"2026-06-05T06:46:15.197025399Z","UpdatedAt":"2026-06-06T19:31:40.473717466Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2606.05382","arxiv_id":"2606.05382","title":"Synthetic Contrastive Reasoning for Multi-Table Q\u0026A","abstract":"Multi-table question answering requires models to retrieve relevant evidence, link schemas, and perform compositional reasoning across relational tables. Existing multi-table Q\u0026A resources typically provide questions and final answers but lack reasoning supervision that explains how answers are derived. To address this gap, we construct a synthetic contrastive reasoning-trace dataset for MMQA by generating validated positive traces and plausible negative traces with heterogeneous LLMs. We then use the resulting preference pairs to fine-tune open-weight LLMs with Contrastive Preference Optimization (CPO). Across Qwen3-14B, Mistral-8B, and Llama-3.1-8B, CPO achieves absolute average improvements over Q\u0026A supervised fine-tuning ranging from 9.7%-16.3%, with gains up to 21 percentage points on MMQA. Ablations show that heterogeneous positive and negative trace generators strengthen the contrastive signal, and automated as well as human evaluations indicate that the generated pairs are largely faithful, coherent, and meaningfully contrastive.","short_abstract":"Multi-table question answering requires models to retrieve relevant evidence, link schemas, and perform compositional reasoning across relational tables. Existing multi-table Q\u0026A resources typically provide questions and final answers but lack reasoning supervision that explains how answers are derived. To address this...","url_abs":"https://arxiv.org/abs/2606.05382","url_pdf":"https://arxiv.org/pdf/2606.05382v1","authors":"[\"Ankit Pratap Singh\",\"Xin Su\",\"Phillip Howard\"]","published":"2026-06-03T19:35:54Z","proceeding":"cs.AI","tasks":"[\"cs.AI\"]","methods":"[\"Large Language Model\"]","has_code":false}