LLM-Powered Automated Research Reporting for Computer Vision Pipelines: From Training Metrics to Scientific Narratives

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Post-training analysis in computer vision produces structured CSV telemetry containing loss trajectories, mAP progressions, and precision-recall dynamics. Synthesizing these metrics into actionable, publication-ready research reports typically demands 30–60 minutes of manual human drafting per experiment. We present a three-stage automated reporting pipeline within the wyoloservice2 ecosystem that (1) parses training logs, (2) executes local narrative synthesis via DeepSeek-V4 through the OpenCode CLI, and (3) compiles branded Markdown and DOCX documents with embedded graphics and MLflow artifact registration. Evaluated across 50 complete YOLO training runs by a blind committee of three independent domain experts, LLM-generated reports achieved competitive performance against expert human drafting across factual accuracy, readability, completeness, and actionability, generating complete documents in 45.2 seconds locally with zero external API data egress.