Published April 15, 2026 | Version 0.1.0

Monolith: Differentiable EML Trees for Symbolic Regression via Gradient Descent on a Universal Operator

Description

Differentiable implementation of EML trees (arXiv 2603.21852) as trainable PyTorch modules for symbolic regression. Demonstrates recovery of 7/7 elementary functions with ≤24 parameters via gradient descent on a single universal operator. Includes hierarchical training for depth 4+, symbolic decompilation, and honest baselines against PySR and MLP.

Files

Monolith.pdf

Files (528.7 kB)

Name Size Download all
md5:850821c0f9e76487e01950c296c86c16
528.7 kB Preview Download

Additional details

Related works

Is supplement to
Publication: arXiv:2603.21852 (arXiv)

Software

Repository URL
https://github.com/seetrex-ai/monolith
Programming language
Python
Development Status
Active

References

  • Odrzywołek, A. (2026). All elementary functions from a single binary operator. arXiv:2603.21852