Auto Researching, not hyperparameter tuning: Convergence Analysis of 10,000 Experiments

TL;DR

This study demonstrates that LLM agents primarily perform genuine architecture search rather than hyperparameter tuning, with architectural choices explaining most of the performance variance in ML experiment design.

Contribution

The paper provides the first large-scale empirical analysis of LLM-guided combinatorial ML experiment design, highlighting the dominance of architecture discovery over hyperparameter tuning.

Findings

Architectural choices explain 94% of performance variance.

LLM agents discover novel effective architectures like V-JEPA2 with Zipformer.

Power-law convergence indicates broad exploration costs, not inefficiency.

Abstract

When LLM agents autonomously design ML experiments, do they perform genuine architecture search -- or do they default to hyperparameter tuning within a narrow region of the design space? We answer this question by analyzing 10,469 experiments executed by two LLM agents (Claude Opus and Gemini 2.5 Pro) across a combinatorial configuration space of 108,000 discrete cells for dashcam collision detection over 27 days. Through ANOVA decomposition, we find that \textbf{architectural choices explain 94\% of performance variance} ($F = 1324$, $\eta^2 = 0.94$), while hyperparameter variation within a fixed architecture explains only 6\%. Cross-task validation on a second collision dataset confirms this finding (75\% architecture-explained variance) with a \emph{different} winning backbone, confirming genuine architecture discovery. The agents' key contribution is discovering that V-JEPA\,2 video features with Zipformer temporal encoders achieve 0.9245 AP -- a configuration no human proposed -- and concentrating search on productive architectural regions: at $N = 50$, LLM-guided search reaches AP $= 0.985$ versus $0.965$ for from-scratch random search. Post-bugfix convergence follows a power law ($c = 0.11$, $R^2 = 0.93$); the low exponent reflects the cost of broad exploration, not inefficiency, since the LLM discovers qualitatively better regions than random or Bayesian baselines. We characterize multi-agent search dynamics via entropy cycles and Jensen--Shannon specialization, providing the first large-scale empirical framework for LLM-guided combinatorial ML experiment design.