Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
FunctionEvolve: Structure-Guided Symbolic Regression with LLMs
TL;DR
Symbolic regression seeks to derive explicit scientific laws from data, but existing methods often lack structural guidance. FunctionEvolve is a new evolutionary framework that uses expression trees to enhance the search process for symbolic recovery.
✦ Why It Matters
Engineers can leverage FunctionEvolve to improve the accuracy of symbolic regression in their data analysis tasks.
Key Takeaways
- Utilize FunctionEvolve for enhanced symbolic regression tasks.
- Incorporate expression trees for better structural representation.
- Focus on local edits to maintain useful mathematical subexpressions.
- Apply structure-aware fitting to improve scoring reliability.
Full Summary
Symbolic regression is a technique used to identify mathematical expressions that describe data, but many current methods are limited by their inability to effectively utilize structural information. FunctionEvolve is introduced as an evolutionary framework that organizes the search for symbolic expressions using expression trees, which represent mathematical functions hierarchically.
This approach incorporates structural summaries to diversify parent selection, allows for local edits to preserve useful subexpressions, and employs structure-aware coefficient fitting to improve scoring reliability. In experiments on the LLM-SRBench benchmark, FunctionEvolve, utilizing the Claude Opus 4.6 model, successfully recovered 107 exact forms, achieving an 82.9% success rate at 50 candidates and a 55.8% success rate at the top candidate, significantly surpassing previous results.
The study also highlights that the visibility of structure during the search process is crucial for reliable recovery, with LLM-guided refinements and optimized coefficient fitting being key components. These findings suggest that incorporating structural awareness can greatly enhance symbolic regression tasks.