Forethought Enables Verifiable AI Reasoning with Neurosymbolic Programs

Current AI agentic workflows often rely on complex reasoning traces within a language model's context, which are difficult to verify and costly to scale. Forethought introduces a novel neurosymbolic reasoning system that addresses these limitations by representing reasoning as an explicit, verifiable program. This program is constructed from a library of both symbolic and neural primitives, composed using a domain-specific language. This approach yields concrete reasoning programs that can be inspected and modified before deployment, offering unprecedented transparency. Evaluated across five benchmarks, Forethought significantly boosted base-model accuracy by approximately 30% relative, outperforming vanilla prompting and other advanced methods. Notably, it enabled smaller, non-reasoning models to compete with or surpass frontier models, requiring three orders of magnitude less post-training investment, while remaining model-agnostic and fully auditable.