FormulaSPIN: Self-Play Fine-Tuning for Natural Language to Spreadsheet Formula Generation

Short summary

FormulaSPIN is a self-play fine-tuning framework for natural language to spreadsheet formula generation that overcomes the limitations of static supervised data. It resolves vanilla SPIN's contradictory gradient problem by using binary executability to separate semantic errors from valid stylistic variants, training via an adaptive curriculum. On NL2FORMULA it achieves 74.9% exact match and 87.1% execution accuracy, matching models trained with additional preference annotations while outperforming both SFT and frontier proprietary models.

• •Self-play framework uses execution feedback to separate semantic errors from stylistic variants in formula generation
• •Achieves 74.9% exact match and 87.1% execution accuracy on NL2FORMULA, outperforming SFT and frontier proprietary models
• •Introduces ExecVote semantic-level voting to handle multiple valid formula formulations

Generated with AI, which can make mistakes.