EvoForest: A Novel Machine-Learning Paradigm via Open-Ended Evolution of Computational Graphs

Researchers Kamer Ali Yuksel and Hassan Sawaf have introduced EvoForest, a novel machine learning paradigm that moves beyond the standard practice of optimizing weights within a fixed model architecture. Instead, EvoForest performs open-ended evolution of computational graphs themselves. This hybrid neuro-symbolic system jointly evolves reusable computational structures, callable function families (like projections and gates), and trainable continuous parameters within a shared directed acyclic graph (DAG). For each candidate graph configuration, it uses a lightweight Ridge regression readout to score the resulting data representation against a non-differentiable, cross-validation-based target. The system then uses structured feedback from this evaluation to guide future mutations, which are powered by a large language model (LLM).

This approach is designed for complex, structured prediction problems where the core challenge is discovering *what* to compute from the data—such as identifying key transformations, invariances, or interaction structures—rather than simply fitting parameters. EvoForest demonstrated its potential by competing in the 2025 ADIA Lab Structural Break Challenge. After 600 evolution steps, the system reached a ROC-AUC score of 94.13%, significantly exceeding the publicly reported winning score of 90.14% achieved under the same evaluation protocol. This result highlights the paradigm's ability to autonomously discover effective computational strategies for tasks where traditional gradient-based optimization is difficult or impossible, opening new avenues for automated machine learning (AutoML) and interpretable model design.