FormulaSPIN Boosts Spreadsheet Formula Generation via Self-Play

Generating spreadsheet formulas from natural language remains a significant challenge for millions of users, with existing methods often limited by static supervised data. This paper introduces FORMULASPIN, a novel self-play framework designed to overcome these limitations and enable iterative self-improvement in formula generation without requiring additional labeled data. Traditional self-play methods struggle with this task because they penalize all non-matching outputs uniformly, even if they are semantically equivalent but stylistically different. FORMULASPIN addresses this by exploiting the unique characteristic of formula generation: binary executability. This allows the system to differentiate between semantic errors and valid stylistic variations, providing more nuanced feedback. The framework frames training as a two-player game where the model learns to prefer ground-truth formulas over its previous versions, with execution feedback guiding an adaptive curriculum from semantic correctness to stylistic refinement. Additionally, ExecVote, a semantic-level voting mechanism, further enhances accuracy by handling multiple valid formula formulations. FORMULASPIN achieves state-of-the-art performance on multiple benchmarks, demonstrating the potential of self-play for data-scarce executable domains.