Agri-CPJ: Training-Free Framework for Agricultural Pest Diagnosis Using LLM-as-a-Judge

The Agri-CPJ (Caption-Prompt-Judge) framework has been introduced to tackle two persistent issues in diagnosing crop diseases through field images: the generation of incorrect species names and unclear reasoning. This training-free few-shot approach employs a substantial vision-language model to create an initial structured morphological caption, which is then refined through multi-dimensional quality gating before any diagnostic queries are addressed. It generates two potential responses from different perspectives, with an LLM judge determining the more robust option based on specific domain criteria. Ablation studies reveal that caption refinement significantly impacts accuracy, as omitting it leads to consistent performance declines in both models evaluated. On CDDMBench, combining GPT-5-Nano with captions from GPT-5-mini shows enhanced results. The research can be found on arXiv under identifier 2604.23701.