Using Probabilistic Programs to Train Inductive Reasoning in Large Language Models

Researchers introduce Program-based Posterior Training (PPT), a novel fine-tuning method that uses probabilistic programs to train LLMs on inductive reasoning tasks. By generating synthetic scenarios and using probabilistic inference to create distributional targets, the approach significantly improves model accuracy on uncertainty estimation while better aligning with human judgment.

This research addresses a critical gap in LLM post-training: while existing methods excel at deductive reasoning (math, coding), they struggle with inductive tasks requiring uncertainty quantification from incomplete information. PPT represents a methodological advance by circumventing two major obstacles—the scarcity of high-quality labeled datasets and the difficulty of training on inherently probabilistic targets.

The approach leverages LLMs' generative capabilities to create 10,000 diverse scenarios encoded as probabilistic programs, then uses formal inference methods to compute correct distributional responses. Fine-tuning against these soft labels teaches models to internalize uncertainty rather than simply rescaling outputs through temperature adjustments. This distinction matters: improvements persist even when controlling for post-hoc calibration tricks, indicating genuine model improvement.

For the AI development community, PPT demonstrates that synthetic data generation via probabilistic programming can overcome dataset curation bottlenecks while maintaining quality. The method's transferability to external benchmarks and human-labeled evaluations suggests broader applicability beyond academic settings. This is particularly valuable for domains like risk assessment, scientific reasoning, and decision-support systems where confidence estimates matter as much as point predictions.

Looking forward, the scalability of this approach remains unclear—whether performance continues improving beyond 10,000 scenarios and how computational costs scale with model size. Integration with deployed LLM systems could enable more reliable uncertainty quantification in real-world applications, though the probabilistic programming requirement may limit adoption by practitioners lacking formal inference expertise.

• →PPT uses LLM-generated probabilistic programs to create high-quality training targets for inductive reasoning tasks
• →The method achieves genuine improvements in calibration that aren't explained by simple output rescaling techniques
• →Synthetic scenario generation addresses dataset scarcity challenges in training uncertain reasoning capabilities
• →Results transfer to external benchmarks and show improved alignment with human judgment on estimation tasks
• →The approach treats uncertainty as a learnable capability rather than a post-processing artifact