Reinforcing Numerical Reasoning in LLMs for Tabular Prediction via Structural Priors

Researchers propose PRPO (Permutation Relative Policy Optimization), a reinforcement learning framework that enhances large language models' numerical reasoning capabilities for tabular data prediction. The method achieves performance comparable to supervised baselines while excelling in zero-shot scenarios, with an 8B parameter model outperforming much larger models by up to 53.17%.

• →PRPO uses column-permutation invariance as a structural prior to improve LLM performance on tabular data prediction tasks.
• →The framework transforms sparse rewards into dense signals, enabling better numerical reasoning with limited supervision.
• →The method matches fully supervised baselines while dominating in zero-shot settings and performing on par with 32-shot strong baselines.
• →An 8B parameter model significantly outperforms DeepSeek-R1 (685B parameters) by up to 53.17%, demonstrating remarkable efficiency.
• →The approach bridges the gap between traditional tabular prediction methods and reasoning-capable LLMs with improved interpretability.