SymStep Boosts LLM Logical Reasoning to 100% on Key Benchmarks

A novel approach named SymStep significantly enhances the capabilities of large language models (LLMs) in handling logical reasoning tasks that are rich in constraints. This technique compels an LLM to assert one atomic statement at a time (for instance, "DEDUCE: Alice, pet, Cat"), followed by a lightweight constraint propagator that verifies the statement's consistency with previously accepted deductions, discarding contradictions and automatically inferring implied facts. An upgraded version, SymStep+G, offers MRV (Minimum Remaining Values) guidance after each accepted claim, steering the LLM towards the most constrained unresolved variable. In tests on a 35-puzzle subset of ZebraLogicBench, SymStep+G achieved 97%, while both Direct prompting and Chain-of-Thought (CoT) scored 0%. SymStep attained 100% on AR-LSAT analytical reasoning tasks compared to CoT's 87%. On LGP-14, SymStep+G also scored 100%, while CoT and Logic-LM, the previous strongest symbolic+LLM baseline, both scored 0%. Ablation studies indicate that MRV guidance is essential for optimal performance. This method resolves the shortcomings of CoT prompting in constraint-heavy tasks, where unverified errors can accumulate unnoticed. The research is accessible on arXiv with ID 2607.23055.