Vision-language models see more about agriculture than they report, and test-time scaling remains bounded by parametric knowledge
The study tested if Vision-Language Models (VLMs) can use test-time reasoning to improve zero-shot agricultural diagnosis without retraining. While simple scaling failed, structuring reasoning around a fixed diagnostic rubric and using a candidate generation method significantly boosted performance. This shows VLMs possess hidden agricultural knowledge that can be unlocked through better prompting and structured verification.