Research map / Vision-language and multimodal / Multimodal LLMs

Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs

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.

Introduces Geo-CoT380k, a 384k-sample structured chain-of-thought dataset for remote sensing, and RSThinker, a GLM-4.1V-9B model trained with SFT then GRPO to produce verifiable grounded reasoning traces.

Problem

Remote sensing VLMs map imagery end-to-end to answers, leaving reasoning unverifiable and prone to hallucination on complex analytical tasks such as counting, grounding, detection, captioning and VQA.

Approach

Model remote sensing analysis as a Perceptually-Grounded Geo-CoT (planning, grounding, synthesis) with rationales retrofitted from ground-truth annotations via GPT-4V under constrained prompting, instilled by SFT and refined by GRPO with task-specific metric-based rewards.

Key results

• RSThinker reaches 96.89 on RESISC45 and 98.17 on AID scene classification, outperforming the listed baselines.
• On VRSBench-VQA Existence, RSThinker scores 92.36 versus 88.89 for ChatGPT-5 and 87.50 for Kimi-VL-Thinking.
• On RSICD captioning RSThinker achieves CIDEr 99.83 versus 85.82 for EarthDial.

Surprising findings

• Applying GRPO without the Geo-CoT SFT stage is insufficient to instill the reasoning scaffold, per the ablation.
• Without KL regularization GRPO causes a collapse of the learned reasoning format.

Limitations

• Rationales generated by GPT-4V may inherit its stylistic biases and are not independently verified for faithfulness.
• Failure case where the model misidentifies a dock extension as a ship, showing textual verification can act as a stylistic heuristic.
• Some baseline numbers (e.g., Claude, ChatGPT-5) are reported without detailed protocol; RSThinker does not win every column (e.g., VRSBench-VQA Quantity).

Experimental setup

Datasets: VRSBench-train FIT-RS NWPU-RESISC45 AID DIOR-RSVG DOTAv2 HRRSD RSVQA-HR NWPU-Captions RSICD RSTMD

Benchmarks: VRSBench-VQA RSVQA-HR RESISC45 AID WHU-RS19 SIRI-WHU UCMerced RSITMD NWPU-Captions RSICD VRSBench-Cap UCM-Captions

Metrics: accuracy mAP IoU MAE BLEU-4 METEOR CIDEr ROUGE-L

Baselines: ChatGPT-5 Gemini-2.0-flash Claude-sonnet-4 Qwen2.5-VL MiniGPT-v2 Kimi-VL-Thinking GLM-4.1V-Thinking VHM SkySenseGPT EarthDial

Builds on: GRPO Chain-of-Thought GLM-4.1V Visual CoT

Closest papers by meaning

• RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.92)
• GeoCoT: Towards Reliable Remote Sensing Reasoning with Manifold Perspective (CVPR 2026 · similarity 0.92)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.92)
• TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.91)
• Asking like Socrates: Socrates helps VLMs understand remote sensing images (CVPR 2026 · similarity 0.91)
• CHOICE: Benchmarking the Remote Sensing Capabilities of Large Vision-Language Models (NeurIPS 2025 · similarity 0.91)