YUKTI: Uncertainty-Aware Optimization from Natural Language with Regret Bounds and Assumption-Robust Pareto Frontiers

Short summary

YUKTI reframes LLM-based optimization by replacing point-valued assumptions with a typed-proposition graph carrying coefficient uncertainty and provenance. It introduces Assumption-Robust Pareto Frontiers that resample assumptions to score action survival rates, with a proven regret bound. Validation shows the robust compromise cuts mean and tail regret by over 90% versus naive point plans, and an LLM given correct numbers incurs ~47x the held-out regret of YUKTI — demonstrating LLMs are formulators, not solvers.

• •YUKTI replaces point-valued coefficients with uncertainty-typed proposition graphs and Assumption-Robust Pareto Frontiers
• •Proven regret bound makes survival rate (rho) an exact factor of decision regret; cuts regret by 90%+ vs naive plans
• •LLMs with correct numbers incur ~47x the regret of YUKTI, confirming LLMs are formulators, not solvers

Generated with AI, which can make mistakes.