YUKTI cuts decision regret by 90% by modeling uncertainty in language-based planning

Current AI pipelines (NL4Opt, OptiMUS, ORLM) for converting natural language into decision plans make a critical mistake: they commit to a single objective and point-valued coefficients, then solve once. This fragile approach assumes every number is exact, leading to plans that fail when assumptions are slightly off. YUKTI, introduced by Suyash Mishra in a new arXiv paper, changes the game by treating every numeric assumption as uncertain. Its core representation is a typed-proposition graph where relationships carry shape priors, coefficient uncertainty, and provenance. The system routes each decision stage to an appropriate solver (exact, nonlinear, or evolutionary) and couples stages via a distributional Pareto hand-off, introducing Assumption-Robust Pareto Frontiers (ARPF). ARPF resamples assumptions—including structural epsilon-contamination—to score how often each action survives (rho), and Mishra proves a bound making rho an exact factor of decision regret.

Validation across three scenarios shows dramatic improvements. Under controlled misspecification, YUKTI's robust compromise cuts mean and tail regret by over 90% versus a naive point plan. On a regulated commercial decision, it optimizes inside a lawful action space and prices downside risk in euros. On a real public dataset of 41,188 decisions, an out-of-sample backtest beats the logged status quo by 34% and a naive point rule by 4% while reducing the optimizer's curse. In a head-to-head, an LLM given correct numbers and single-objective optimization both incur about 47x the held-out regret of YUKTI. The paper also addresses long-range causal coupling, identifying where forward hand-off becomes unsound and must become backward-induction causal policy. YUKTI represents a fundamental shift: LLMs are formulators, not solvers, and robust decision-making requires embracing uncertainty.