Type: url
Original: https://arxiv.org/pdf/2601.09745
Status: generated
Summary
An IBM Research case study (January 2026) arguing that natural language augmented with lightweight mathematical notation and written in LaTeX can serve as an intermediate specification language between informal requirements and heavyweight formal methods such as Z. Applying AI-driven review and refinement of the specification before any code is generated — rather than correcting the code from its execution results — the authors report roughly a sixth of the development effort and a correct implementation on the first code-generation attempt for a nontrivial organizational knowledge-growth simulation. A final experiment deliberately summarizes the precise specification into vague prose to show how ambiguity blurs the line between the business logic an analyst must control and the peripheral aspects a model may freely decide.