WeedExpert-R1: Incentivizing Botanical Reasoning in MLLMs with Reinforcement Learning for Precision Weed Grounding

2026-07-17

AI Summary

This study addresses the challenge of achieving species-level weed identification and instance-level localization in complex agricultural scenes—a task where existing methods suffer from limited interpretability and poor cross-regional generalization. To overcome these limitations, the authors propose WeedExpert-R1-4B, a novel framework that integrates botanical knowledge with multimodal large language models. The approach leverages a manually curated plant trait dictionary, an Auditor-Synthesizer LLM workflow, supervised fine-tuning, and a newly introduced Group Relative Policy Optimization reinforcement learning algorithm, which incorporates a verifiable reward mechanism and a domain-specific chain-of-thought synthesis pipeline. Evaluated on 37 weed species, the method achieves 75.82% exact-set accuracy at IoU=0.5, substantially outperforming closed-source models such as GPT-5.4 and Gemini-3.1-Pro, as well as larger open-source alternatives, while demonstrating strong generalization to unseen species.

Abstract

Precision weed control requires species-level identification and instance-level localization. However, conventional object detectors use a closed vocabulary, limiting their deployment across regions, and cannot explain their predictions in complex agricultural scenes. Multimodal large language models (MLLMs) offer visual grounding and reasoning capabilities, but insufficient botanical knowledge can cause hallucinations in fine-grained weed identification. This study introduces WeedExpert-R1, a multimodal model that learns visually grounded botanical reasoning through verifiable rewards. A domain-specific Chain-of-Thought synthesis pipeline combines a human-curated botanical trait dictionary with an Auditor-Synthesizer LLM workflow to generate reasoning data for supervised fine-tuning. Group Relative Policy Optimization is then applied with rewards for format, accuracy, instance count, and response length. Across 37 weed species from six datasets, WeedExpert-R1-4B achieved 75.82 percent exact-set precision at an IoU threshold of 0.5, 89.30 percent precision, and 87.81 percent recall. It outperformed proprietary models, including GPT-5.4 and Gemini-3.1-Pro, and larger open-source models, including Qwen3-VL-30B-Instruct and Gemma-4-31B-it. Results on unseen species further demonstrate its open-vocabulary capability and potential for deployment across diverse regions and crops without retraining.

Problem

• precision weed control
• fine-grained weed identification
• multimodal large language models
• botanical reasoning
• open-vocabulary grounding

Innovation

• Multimodal Large Language Models
• Reinforcement Learning
• Botanical Reasoning
• Visual Grounding
• Open-Vocabulary Object Detection