Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Agri-R1: Empowering Generalizable Agricultural Reasoning in Vision-Language Models with Reinforcement Learning

Wentao Zhang, Lifei Wang, Lina Lu, MingKun Xu, Shangyang Li, Yanchao Yang, Tao Fang

TL;DR

Agri-R1 addresses data efficiency, interpretability, and cross-domain robustness in agricultural vision-language diagnosis by combining automated reasoning data synthesis with GRPO-based reinforcement learning. A two-stage pipeline first generates reasoning exemplars using DeepSeek-VL2 and filters them with GPT-4, then trains a 3B model with a domain-aware, three-component reward under group-relative policy optimization. Results on CDDMBench show substantial gains in disease recognition (+23.2% relative), knowledge QA (+33.3%), and cross-domain generalization (+26.1 points), with ablations confirming the synergy between reasoning data and GRPO. The approach eliminates manual CoT annotation and delivers data-efficient, interpretable reasoning that remains robust under domain shifts, enabling practical deployment in resource-constrained agricultural settings.

Abstract

Agricultural disease diagnosis challenges VLMs, as conventional fine-tuning requires extensive labels, lacks interpretability, and generalizes poorly. While reasoning improves model robustness, existing methods rely on costly expert annotations and rarely address the open-ended, diverse nature of agricultural queries. To address these limitations, we propose \textbf{Agri-R1}, a reasoning-enhanced large model for agriculture. Our framework automates high-quality reasoning data generation via vision-language synthesis and LLM-based filtering, using only 19\% of available samples. Training employs Group Relative Policy Optimization (GRPO) with a novel proposed reward function that integrates domain-specific lexicons and fuzzy matching to assess both correctness and linguistic flexibility in open-ended responses. Evaluated on CDDMBench, our resulting 3B-parameter model achieves performance competitive with 7B- to 13B-parameter baselines, showing a +23.2\% relative gain in disease recognition accuracy, +33.3\% in agricultural knowledge QA, and a +26.10-point improvement in cross-domain generalization over standard fine-tuning. Ablation studies confirm that the synergy between structured reasoning data and GRPO-driven exploration underpins these gains, with benefits scaling as question complexity increases.

Agri-R1: Empowering Generalizable Agricultural Reasoning in Vision-Language Models with Reinforcement Learning

TL;DR

Agri-R1 addresses data efficiency, interpretability, and cross-domain robustness in agricultural vision-language diagnosis by combining automated reasoning data synthesis with GRPO-based reinforcement learning. A two-stage pipeline first generates reasoning exemplars using DeepSeek-VL2 and filters them with GPT-4, then trains a 3B model with a domain-aware, three-component reward under group-relative policy optimization. Results on CDDMBench show substantial gains in disease recognition (+23.2% relative), knowledge QA (+33.3%), and cross-domain generalization (+26.1 points), with ablations confirming the synergy between reasoning data and GRPO. The approach eliminates manual CoT annotation and delivers data-efficient, interpretable reasoning that remains robust under domain shifts, enabling practical deployment in resource-constrained agricultural settings.

Abstract

Agricultural disease diagnosis challenges VLMs, as conventional fine-tuning requires extensive labels, lacks interpretability, and generalizes poorly. While reasoning improves model robustness, existing methods rely on costly expert annotations and rarely address the open-ended, diverse nature of agricultural queries. To address these limitations, we propose \textbf{Agri-R1}, a reasoning-enhanced large model for agriculture. Our framework automates high-quality reasoning data generation via vision-language synthesis and LLM-based filtering, using only 19\% of available samples. Training employs Group Relative Policy Optimization (GRPO) with a novel proposed reward function that integrates domain-specific lexicons and fuzzy matching to assess both correctness and linguistic flexibility in open-ended responses. Evaluated on CDDMBench, our resulting 3B-parameter model achieves performance competitive with 7B- to 13B-parameter baselines, showing a +23.2\% relative gain in disease recognition accuracy, +33.3\% in agricultural knowledge QA, and a +26.10-point improvement in cross-domain generalization over standard fine-tuning. Ablation studies confirm that the synergy between structured reasoning data and GRPO-driven exploration underpins these gains, with benefits scaling as question complexity increases.
Paper Structure (54 sections, 8 equations, 5 figures, 9 tables)

This paper contains 54 sections, 8 equations, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Agricultural Vision-Language Models
  4. Chain-of-Thought for Interpretability
  5. Reinforcement Learning for VLM Post-Training
  6. Methodology
  7. Generative Reasoning Enhancement
  8. GRPO Reinforcement Learning
  9. Group Relative Policy Optimization
  10. Reward Function Design
  11. Domain Vocabularies.
  12. Format Reward.
  13. Answer Keyword Reward.
  14. Reasoning Quality Reward.
  15. Dynamic Evaluation.
  16. ...and 39 more sections

Figures (5)

  • Figure 1: Overview of our Two-Stage GRPO Framework for Agricultural Disease Reasoning. Stage 1 transforms raw VQA pairs into reasoning exemplars: DeepSeek-VL2 generates reasoning chains, GPT-4 filters outputs (threshold $\tau$=8.0/10.0). Stage 2 employs GRPO-based policy learning with Domain Vocabulary Construction: 5-tier fuzzy matching handles linguistic diversity, three-component reward function (Format + Answer + Reasoning) guides optimization, and group relative advantage normalization (n=3 samples) enables stable updates. This pipeline enables our 3B model to learn robust reasoning from synthesized data.
  • Figure 2: AgMMU task distribution. Reasoning-Enhanced GRPO (red) outperforms SFT (blue) on visual tasks while showing balanced performance.
  • Figure 3: Disease recognition improvement distribution. Point size represents training data proportion. Green points show >15% gains, red points <-15% declines.
  • Figure 4: Ablation study on Disease Knowledge QA (20 samples). GRPO (no explicit reasoning, purple) provides baseline gains (+4% to +28%), while Reasoning-Enhanced GRPO (red) achieves 2.2× amplification on complex reasoning (+61%).
  • Figure 5: A comparison of diagnostic reasoning. Our Reasoning-Enhanced GRPO (top) produces structured explanations with actionable details, while standard GRPO (bottom) provides minimal operational guidance.