A new AI reasoning system swaps buried chain-of-thought for programs you can actually open and edit before they run.
Forethought, introduced in a preprint from arXiv, reframes reasoning not as a learned behavior baked into model weights but as an explicit program built from a library of symbolic and neural building blocks. Those building blocks are composed through a domain-specific language, producing what the researchers call reasoning programs — concrete artifacts that can be inspected and modified before deployment. Evaluated across five benchmarks, Forethought improved base-model accuracy by roughly 30% relative, outperforming vanilla prompting, reinforcement learning scaffolds, and prompt-evolution methods. The kicker: a standard non-reasoning model with Forethought bolted on reportedly matches a dedicated reasoning model while requiring about three orders of magnitude less post-training compute.
That last claim cuts at the core tension in AI reasoning research. Test-time scaling — training models to search over long chains of thought — has become the dominant approach, but it hides its work inside weights, resists auditing, and is expensive to run. Forethought's pitch is that verifiability and auditability are not just nice-to-have properties; they are what you need before you trust an agentic system to call real tools with real consequences.
Small models matching frontier models is a pitch the field has heard before, and benchmark gains have a way of shrinking on contact with production. But the auditable-program angle is harder to dismiss — and notably absent from every major lab's current reasoning roadmap.