Precision weed control requires both species-level identification and per-instance localization, butconventional object detectors are constrained by a closed-vocabulary paradigm that hinders cross-regiondeployment and lack the reasoning capability to justify their predictions in complex agricultural scenes.Recently, multimodal large language models (MLLMs) have demonstrated strong visual perceptionand reasoning in visual grounding, offering a promising alternative. However, their insufficient domainspecificbotanical knowledge often leads to hallucinations during fine-grained weed identification. Inthis study,WeedExpert-R1 is introduced, a multimodal reasoning model that, following the R1-styletraining paradigm, incentivizes visually grounded botanical reasoning through verifiable rewards.Specifically, a domain-specific Chain-of-Thought (CoT) synthesis pipeline is proposed. It pairs ahuman-curated botanical trait dictionary, covering leaf shape, margin, petiole, and stem morphology,with an Auditor–Synthesizer LLM workflow to generate high-quality reasoning data for supervisedfine-tuning as a cold start. This pipeline bridges the gap between textual botanical knowledge and thevisual perception of MLLMs. Group Relative Policy Optimization (GRPO) is then applied withverifiable rewards (format, accuracy, count, and length penalty) to further enhance the model’scapability to perceive and localize diverse weeds. WeedExpert-R1-4B achieves 75.82%, 89.30%,and 87.81% on Precision@(F1=1, IoU≥0.5), Precision@0.5, and Recall@0.5 across 37 weed specieson a benchmark suite of six weed datasets (3SeasonWeedDet10, CottonWeedDet12, CottonWeedDet3,Weed-crop,Weed25, and PREEC), and outperforms both frontier proprietary models (GPT-5.4, Gemini-3.1-Pro) and larger open-source baselines (Qwen3-VL-30B-Instruct, Gemma-4-31B-it). Moreover,qualitative results on unseen species further illustrate the open-vocabulary capability ofWeedExpert-R1,suggesting potential for deployment across diverse regions and crops without retraining.