Generating Domain Specific Natural Language SAT Reasoning Datasets
Published in NeurIPS Workshop on Efficient Reasoning (2025), 2025
Recommended citation: Patra, S., Ramani, K., Borrajo, D., & Gopalakrishnan, S. (2025). Generating domain specific natural language SAT reasoning datasets. In NeurIPS Workshop on Efficient Reasoning.
Abstract
Evaluating the multi-hop logical deductions of LLMs requires rigorous ground-truth reasoning datasets. We present a programmatic framework to translate complex Boolean satisfiability (SAT) instances into natural language narratives with precise logical dependencies, creating un-gameable evaluation datasets for multi-step reasoning and deduction.
Recommended citation: Patra, S., Ramani, K., Borrajo, D., & Gopalakrishnan, S. (2025). Generating domain specific natural language SAT reasoning datasets. In NeurIPS Workshop on Efficient Reasoning.