Forethought: Verifiable Reasoning from Neurosymbolic Primitive Programming

Forethought proposes a neurosymbolic reasoning system that represents agent reasoning as explicit, executable programs composed from symbolic and neural primitives in a domain-specific language. Unlike long chain-of-thought traces embedded in model context, these programs can be inspected and modified before deployment. The authors report evaluations on five benchmarks, with roughly a 30% relative accuracy improvement over base models. They also claim that a non-reasoning model augmented with Forethought can compete with a dedicated reasoning model while requiring about three orders of magnitude less post-training investment. The abstract does not identify the benchmarks or provide detailed results.