FormulaGPT is an innovative approach to symbolic regression, leveraging the strengths of reinforcement learning (RL) algorithms and the efficiency of transformer-based models.

• Sparse Reward Learning Data: Utilizes the historical data of RL-based symbolic regression as training data for a GPT.
• Distilled Knowledge: The RL is distilled into a transformer, enabling a GPT to generate an ‘RL process’ in-context.
• State-of-the-Art Performance: Achieves superior fitting ability on various datasets, including SRBench.
• Efficiency and Robustness: Offers improved results in terms of noise robustness, versatility, and inference speed.

FormulaGPT’s novel application represents a unique intersection of different AI methodologies—combining the high resilience of RL with the rapid inference of transformers. It suggests a compelling research direction for more complex symbolic reasoning tasks and the creation of more efficient AI-driven scientific discovery methods.