Accelerating Hyperparameter Tuning for YOLO-Based Endangered Sea Turtle Classification via an LLM-Agent Recommendation

Hyperparameter tuning for object detection is a time-consuming cognitive task that ML researchers face repeatedly: every candidate configuration demands a full training run, and conventional search algorithms (grid, random, genetic, Bayesian) routinely consume hundreds to thousands of hours per tuned model. We share a simple AI-assisted workflow that collapses this inner loop into a one-shot literature-driven recommendation from an LLM-based research agent (Nexus Science): given a task description, dataset properties, and a target model variant, the agent surveys ≥500 published papers and returns a single hyperparameter configuration in under one hour. Applied to six YOLO variants (v8x, v9e, v10x, 11x, 12x, and 26x) on a seven-species sea turtle classification task, the workflow yields an average mAP@0.5 improvement of +0.089 over Ultralytics defaults—larger than the gain from any architectural upgrade within the same family. Each per-model recommendation completes in 18–42 minutes, eliminating over 99% of the wall-clock cost reported for random-search and GA-based HPO of YOLO models in the literature. We release the full 13-step prompt, recommended configurations, and verification scripts so that other practitioners can reproduce or adapt the methodology with their preferred LLM-based research agent. Together, this work offers a concrete, reproducible case for accelerating hyperparameter tuning by replacing search-based HPO with a one-shot literature-driven LLM recommendation.