Using Probabilistic Programs to Train Inductive Reasoning in Large Language Models

The paper presents a novel approach called Program-based Posterior Training (PPT) for enhancing inductive reasoning in Large Language Models (LLMs). By generating diverse scenarios as probabilistic programs and fine-tuning LLMs on 10,000 programmatically generated scenarios, the method significantly improves estimation accuracy and alignment with human judgments on inductive tasks, demonstrating that models better internalize uncertainty compared to traditional post-hoc temperature scaling. This approach addresses challenges in standard fine-tuning methods and offers a scalable solution for practitioners focusing on real-world reasoning problems.