YUKTI Improves AI Decision-Making with Robust, Verifiable Plans from Language Models

Traditional AI pipelines for decision-making often convert natural language into a single, fixed plan, assuming precise coefficients and objectives. This approach is fragile because real-world decisions involve inherent uncertainties and assumptions, leading to suboptimal or risky outcomes when these guesses are incorrect. YUKTI addresses this by changing the target of autoformulation. Instead of a single plan, YUKTI creates a typed-proposition graph that explicitly incorporates uncertainty, shape priors, and data provenance. It then uses a multi-stage solving process, including exact, nonlinear, and evolutionary solvers, coupled by distributional Pareto hand-offs. A key innovation is the introduction of Assumption-Robust Pareto Frontiers (ARPF), which resample assumptions to evaluate how often each action remains viable, providing a quantifiable measure of decision regret. The framework has been validated across various scenarios, demonstrating over 90% reduction in mean and tail regret compared to naive point plans under controlled misspecification. On commercial decisions, it optimizes inside legal boundaries and quantifies financial downside. Furthermore, an out-of-sample backtest on a large public dataset showed YUKTI outperforming existing methods by a significant margin, confirming its ability to generate more robust and auditable decisions.