Text Before Vision: Staged Knowledge Injection Matters for Agentic RLVR in Ultra-High-Resolution Remote Sensing Understanding

Multimodal reasoning for ultra-high-resolution (UHR) remote sensing (RS) is usually bottlenecked by visual evidence acquisition: the model necessities localizing tiny task-relevant regions in massive pixel spaces. While Agentic Reinforcement Learning with Verifiable Rewards (RLVR) using zoom-in tools offers a path forward, we find that standard reinforcement learning struggles to navigate these vast visual spaces without structured domain priors. In this paper, we investigate the interplay between post-training paradigms: comparing Cold-start Supervised Fine-Tuning (SFT), RLVR, and Agentic RLVR on the UHR RS benchmark. Our controlled studies yield a counter-intuitive finding: high-quality Earth-science text-only QA is a primary driver of UHR visual reasoning gains. Despite lacking images, domain-specific text injects the concepts, mechanistic explanations, and decision rules necessary to guide visual evidence retrieval. Based on this, we propose a staged knowledge injection recipe: (1) cold-starting with scalable, knowledge-graph-verified Earth-science text QA to instill reasoning structures; and (2) "pre-warming'' on the same hard UHR image–text examples during SFT to stabilize and amplify subsequent tool-based RL. This approach achieves a 60.04\% Pass@1 on XLRS-Bench, significantly outperforming larger general-purpose models (e.g., GPT-5.2, Gemini 3.0 Pro, Intern-S1) and establishing a new state-of-the-art. We provide both the automated data pipeline and the rigorous ablation studies that validate this "Text-Before-Vision'' paradigm. Datasets and code will be released.

Reasoning over ultra-high-resolution remote-sensing images is challenging because the model often needs to find very small, task-relevant regions within extremely large images. Although allowing the model to actively zoom in provides a promising way to collect visual evidence, standard reinforcement learning is difficult to apply effectively without prior Earth-science knowledge. In this work, we compare several post-training strategies for this setting and find a somewhat surprising result: high-quality Earth-science question-answering data in text form is a key factor in improving visual reasoning, even though it contains no images. Such text data provides the concepts, explanations, and decision rules needed to guide the model toward useful visual evidence. Based on this finding, we propose a staged “Text-Before-Vision” training recipe: first injecting Earth-science knowledge through verified text-only QA, and then warming up the model on difficult image-text examples before tool-based reinforcement learning. This approach improves performance on XLRS-Bench and outperforms larger general-purpose multimodal models, suggesting that structured domain knowledge is crucial for reliable reasoning over ultra-high-resolution remote-sensing imagery.