While Vision Language Models (VLMs) show advancing reasoning capabilities, their application in meteorology is constrained by a domain gap and a reasoning faithfulness gap. Mainstream Reinforcement Fine-Tuning (RFT) can induce Self-Contradictory Reasoning (Self-Contra), where the reasoning process contradicts the final answer, which is unacceptable in this high-stakes domain.
To address these challenges, we construct WeatherQA, a multimodal multiple-choice benchmark for meteorology comprising 15,400 entries that cover four themes and seven imaging modality tasks. We propose Logically Consistent Reinforcement Fine-Tuning (LoCo-RFT), which introduces a logical consistency reward to resolve Self-Contra. Based on this paradigm and WeatherQA, we present Weather-R1, the first reasoning VLM with logical faithfulness in meteorology, to the best of our knowledge. Weather-R1 (7B) achieves 52.9% accuracy on WeatherQA, a 9.8 percentage point gain over the baseline model Qwen2.5-VL-7B; it surpasses Supervised Fine-Tuning and RFT baselines, exceeds the original Qwen2.5-VL-32B, and improves out-of-domain ScienceQA performance by 4.98 percentage points.
Response Comparison.
🗂️ Folder Structure
This repository provides model checkpoints organized by training strategy and task:
Each task folder contains HuggingFace-style model files such as config.json,
tokenizer.json, and sharded weights like model-00001-of-00004.safetensors.
🚀 Training and Evaluation
Please refer to our official repository: Weather-R1
