Research map / Vision-language and multimodal / Multimodal LLMs

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

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that high-quality Earth-science text-only QA with CoT, used in cold-start SFT and followed by SFT pre-warming on UHR RS VQA and agentic zoom-in RLVR, lifts XLRS-Bench pass@1 to 60.40%.

Problem

Agentic RLVR with zoom-in tools struggles to localize tiny task-relevant regions in ultra-high-resolution remote sensing imagery; the respective roles of SFT, RLVR and agentic RLVR and of domain data are poorly understood in this setting.

Approach

Injects domain knowledge via text-only Earth-science CoT QA (generated from textbooks and papers, filtered with a textbook-derived knowledge graph) as cold-start SFT, then pre-warms with SFT on the same hard UHR image-text VQA before zoom-in agentic RLVR (built on DeepEyes).

Key results

• Staged recipe (ES text CoT cold-start + SuperRS-VQA pre-warming + tool GRPO) reaches 60.40 pass@1 and 96.25 pass@32 on XLRS-Bench, versus 50.01/82.58 for the DeepEyes-style baseline.
• Removing CoT from Earth-science text QA drops pass@1 by 5.91 points (60.40 to 54.49) but pass@32 by only 0.50.
• The 7B model's 60.40 pass@1 surpasses GPT-5.2 (47.53), Gemini 2.5 Pro (45.2) and Intern-S1 (241B, 55.0) as reported on XLRS-Bench.

Surprising findings

• Text-only Earth-science QA improves UHR visual reasoning more than image-text VQA during cold-start SFT.
• Standard GRPO without tools on the domain data drops pass@1 to 25.31, showing RLVR alone can be unstable without domain priors.

Limitations

• Text-only data gains are measured on a single benchmark (XLRS-Bench) and mostly with Qwen2.5-VL-7B.
• Pass@32 gains plateau with more text data, suggesting diminishing returns on the reasoning boundary.
• Text QA data generation relies on GPT-5 and LLM-based filtering, which may introduce hallucinations despite KG verification.

Experimental setup

Datasets: SuperRS-VQA DeepEyes-47K XLRS-Bench s1.1-R1 Earth-science text QA (textbooks and papers)

Benchmarks: XLRS-Bench

Metrics: pass@1 (macro accuracy) pass@32

Baselines: DeepEyes (GRPO w/ tools) GRPO GeoChat ZoomEarth GeoLLaVA-8K GPT-5.2 Gemini 2.5 Pro Intern-S1 Qwen3-VL-235B-A22B Claude 3.7 Sonnet

Builds on: DeepEyes GRPO LLaMA-Factory LightRAG DeepEyes-47K

Closest papers by meaning

• Asking like Socrates: Socrates helps VLMs understand remote sensing images (CVPR 2026 · similarity 0.91)
• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.89)
• RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.88)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.88)
• GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.87)
• ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks (CVPR 2026 · similarity 0.87)