Beyond Adam: AI-Authored Discovery of Symbolic Optimization Rules

Robert Yang

Keywords: AI authorship, autonomous scientific agents, symbolic optimization, evolutionary algorithms, domain-specific languages (DSL), reproducibility, interpretable machine learning, meta-optimization

Abstract:

Can an AI system act as the sole first author of a scientific paper?
We investigate this question through the Algorithmic Greenhouse, an autonomous framework that evolves symbolic optimization rules. A compact domain-specific language (DSL) spans canonical methods such as SGD, Momentum, and Adam, while enabling novel hybrids. Using an evolutionary loop with mutation and elitism, the agent searches this space on analytic landscapes including Rastrigin, Rosenbrock, and Ackley, and evaluates transfer to synthetic regression.

The discovered rules are simple, interpretable formulas that are competitive with standard baselines. More importantly, the entire research pipeline---from DSL design and evolutionary search through experiments, figure generation, and manuscript drafting---was conducted autonomously by the AI agent. Human collaborators provided only high-level oversight. This end-to-end authorship, rather than incremental optimizer performance, is the central contribution: a demonstration that AI can propose hypotheses, implement algorithms, analyze outcomes, and communicate results in a scientific format.

The modest scope of our experiments reflects compute constraints, but the process generalizes: the same framework could be applied to richer DSLs and higher-dimensional tasks such as neural network training. We argue that the Algorithmic Greenhouse should be viewed as a proof-of-concept for responsible AI-driven science, illustrating both the promise and the limits of autonomous AI authorship.

Submission Number: 145