MOT-SR Adds a Dynamic Pareto Front to LLM Equation Discovery, Validated on Gravitational-Wave Orbital Modeling

arXiv 2607.29561·low signalXBlueskyLinkedInCopy link

LLM-based symbolic regression typically optimizes fitting error alone, which pushes models into premature local optima, and lacks mechanisms for uncovering variable dependencies before generating candidates. MOT-SR pairs a Meta Strategy Generator that selects external analytical tools and synthesizes structural strategies from Pareto-optimal equations with an Equation Generator producing candidates, jointly optimizing accuracy, complexity, and generalization against a dynamic Pareto front in a closed loop. It outperforms existing SR methods across 40 standard tasks, and on extreme mass-ratio inspiral orbital modeling its discovered interpretable correction achieves the lowest trajectory-level integration error on held-out configurations.