CPO boosts multi-table QA by 21 pts on Qwen, Mistral, Llama
Synthetic contrastive reasoning traces lift LLM accuracy by up to 21% on complex table queries.
Multi-table question answering (QA) requires models to retrieve evidence, join schemas, and reason compositionally, but existing datasets lack reasoning supervision. Researchers from the paper (arXiv:2606.05382) address this by constructing a synthetic contrastive reasoning-trace dataset for MMQA. They use heterogeneous LLMs to generate validated positive traces (correct reasoning) and plausible negative traces (incorrect reasoning), creating preference pairs for fine-tuning.
Using Contrastive Preference Optimization (CPO), they fine-tune open-weight models Qwen3-14B, Mistral-8B, and Llama-3.1-8B. CPO yields absolute average improvements of 9.7%-16.3% over Q&A supervised fine-tuning, with a maximum gain of 21 percentage points on MMQA. Ablations show that heterogeneous positive and negative generators produce stronger contrastive signals, and both automated and human evaluations confirm the traces are faithful, coherent, and meaningfully contrastive.
- Synthetic dataset of positive and negative reasoning traces for multi-table QA (MMQA) generated by heterogeneous LLMs.
- Contrastive Preference Optimization (CPO) fine-tuning applied to Qwen3-14B, Mistral-8B, and Llama-3.1-8B.
- Achieved 9.7%-16.3% absolute average accuracy improvements over SFT, with gains up to 21 percentage points.
Why It Matters
Enables more reliable compositional reasoning across relational tables for enterprise data analysis and database QA.