Large-scale agricultural grounding data can now be generated automatically, but how training on such data affects a vision-language model's localization, recognition, and refusal capabilities remains unclear. To study these effects, we introduce AgroGround, containing 794,850 automatically grounded agricultural instructions for training and a separate benchmark of 1,478 human-verified images for evaluation. Among these images, 500 are independently re-annotated, yielding an inter-annotator instance F1 of 0.789. We use controlled comparisons to isolate the effects of different training choices on localization, recognition, their joint correctness, and refusal. Grounding-only fine-tuning reduces four-way recognition from 51.7% to 29.2%, with predictions showing a strong candidate-position bias. However, changing only the instruction format raises recognition to 72.7% while maintaining similar grounding performance. Training with healthy-image negatives achieves 94.1% refusal on healthy images, but only 28.6% refusal of incorrect target names on positive images, showing that these two refusal behaviors do not necessarily improve together. Reinforcement learning reveals a further trade-off, improving localization by 0.014 F1 over a matched supervised control while reducing recognition and refusal. Across model families and held-out training sources, localization transfers more reliably than recognition and target-specific refusal. External benchmarks further show that the gains depend on the domain and task. Together, these findings show that grounding fine-tuning changes the balance between localization, recognition, and refusal rather than improving all three uniformly.
翻译:暂无翻译