PestLife: A Lifecycle-Aware Evaluation Benchmark for Multimodal Large Models in Rice Pest Management

PestLife is a multimodal large language model evaluation benchmark designed for rice pest and disease management scenarios, aiming to fill the gap in lifecycle-aware evaluation. Its core is a three-level evaluation framework that systematically tests models' capabilities in species identification, growth stage recognition, and knowledge application, helping to locate model weaknesses and support precision agricultural decision-making.

Traditional pest and disease diagnosis is often treated as a single task, but actual management requires integrating multi-dimensional information—pest damage levels and control strategies vary significantly across different growth stages (e.g., larvae and adults require different approaches). Most existing multimodal evaluation benchmarks ignore the importance of lifecycle awareness, and current models show significant performance differences in fine-grained agricultural tasks (especially growth stage recognition).

PestLife's three-level evaluation framework is divided into:

1. Single-capability evaluation: Species identification (S), growth stage recognition (T), knowledge application (K);
2. Combined reasoning evaluation: Paired tasks such as species-stage joint identification (S-T), species-knowledge application (S-K), stage-knowledge application (T-K);
3. End-to-end comprehensive reasoning: Simultaneously completing species identification, stage judgment, and control recommendation generation (S-T-K). This design can accurately locate weaknesses in the model's reasoning process and provide directions for improvement.

The PestLife dataset is built following strict standards:

• Sources: iNaturalist platform, professional agricultural websites, manually selected images;
• Content: 35 major rice pest species, covering growth stages such as eggs, larvae, pupae, and adults;
• Quality control: Clustering and deduplication, expert annotation verification, generation of 12305 question-answer pairs, multi-stage filtering (blurry/low-quality samples, text-dependent samples), expert consistency verification (≥90%);
• Final scale: 1195 high-quality expert-annotated images.

Zero-shot evaluation of 39 advanced models (32 multimodal large language models +7 baseline models) reveals:

1. Growth stage recognition is a key bottleneck, with higher difficulty than species identification;
2. Quantified the contribution of lifecycle awareness through controlled experiments (performance gap between models with and without stage information);
3. Current models show significant performance differences in fine-grained agricultural tasks, leaving large room for improvement.

Practical value of PestLife: Helps improve models, supports farmers' precision control decisions, reduces pesticide use, and increases crop yields. For continuous expansion, the team developed a web-based crowdsourcing data collection system that allows practitioners to upload real pest images and structured annotations (species, stage). After expert review, these are included in future datasets to ensure the data reflects the diversity of actual agricultural scenarios.

Limitations and future directions of PestLife:

• The dataset focuses on rice pests and needs to be expanded to other crops;
• The 35 species cover major pests, but more edge cases need to be included;
• Current evaluation is based on zero-shot settings; future work will explore the impact of few-shot learning and fine-tuning;
• Apply the evaluation method to other agricultural pest and disease management scenarios.