The Modality Paradox in Autonomous LLM Engineering: Asymmetric Agent Loops and Mathematical Halting.

As Large Language Models (LLMs) are deployed in autonomous closed-loop engineering workflows, a recurring failure mode emerges: the Autonomous Sunk-Cost Fallacy. A mono lithic agent often continues low-yield strategy edits for dozens of iterations despite negligible progress, consuming large compute budgets.This paper focuses exclusively on that failure mode in machine learning engineering loops. We evaluate single-agent and asymmetric Reviewer→Coder architectures across Tabular, Text, and Vision modalities in AEOS. We formalize Sunk-Cost Episodes (SCEs), define theCognitive Agentic Diversity Score (CADS), and introduce the Ω Cognitive Yield Engine for
mathematical stopping. In addition to the baseline modality experiments, we run a targeted mathematical selfreflection ablation (N=36): 3 datasets × 4 architectures × 3 repeats. The ablation shows strong autonomous stopping behavior (35/36 runs halt via reviewer directive) while revealing a modality-dependent optimum: the best architecture is not globally fixed. We conclude that robust autonomous optimization requires both role asymmetry and mathematically
grounded halting; the mathematical gate achieves 35/36 reviewer-issued stops while preserving near-equivalent peak accuracy, with task modality governing the persistence-efficiency tradeoff.