Research map / Vision-language and multimodal / Multimodal LLMs

RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 4 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.

Proposes RemoteReasoner, an RL-trained (GRPO + LoRA) Qwen2.5-VL-7B workflow that outputs boxes for region-level reasoning and derives pixel masks (via SAM2) and contours from them, for multi-granularity geospatial reasoning without task-specific decoders.

Problem

Existing remote sensing reasoning segmentation methods rely on SFT with annotated reasoning traces, are limited to single-task outputs with task-specific decoders, and generalize poorly to out-of-distribution tasks and categories.

Approach

Train an MLLM with GRPO and a composite IoU/count/format reward to output bounding boxes with reasoning, then transform the boxes at inference into pixel masks via SAM2 and into object contours via morphological boundary extraction, yielding three granularities from one forward pass.

Key results

• On EarthReason test pixel reasoning, RemoteReasoner reaches 69.13 cIoU versus 68.25 for SegEarth-R1, without pixel-level training labels.
• On region reasoning (EarthReason val), Acc@0.5 of 66.51 versus 41.21 for Qwen2.5-VL-7B.
• Contour reasoning F1@1 on test is 0.44 versus 0.10 for EPOC.

Surprising findings

• Text-based polygon mask generation performed far worse than box-to-SAM mask generation (cIoU 18.74 with RL vs 69.13).

Limitations

• Weak at detecting small objects such as cars.
• Relies on SAM2 and morphological operations for pixel and contour outputs, so segmentation quality is bounded by box accuracy.
• Only 100 manually selected images across 10 unseen categories used for OOD evaluation.

Experimental setup

Datasets: EarthReason RSVG RRSISD RemoteSAM-270K

Benchmarks: EarthReason RSVG RRSISD RemoteSAM-270K

Metrics: gIoU cIoU Acc@0.5 F1@1 F1@3 Recall@1 Recall@3 ASD HD

Baselines: LISA PixelLM PSALM SegEarth-R1 DeepSeek-VL2-tiny GeoChat Qwen2.5-VL-7B EPOC MGVLF Falcon SkyEyeGPT LHRS-Bot

Builds on: GRPO Qwen2.5-VL SegEarth-R1 RemoteSAM SAM2

Closest papers by meaning

• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.92)
• Asking like Socrates: Socrates helps VLMs understand remote sensing images (CVPR 2026 · similarity 0.90)
• GeoCoT: Towards Reliable Remote Sensing Reasoning with Manifold Perspective (CVPR 2026 · similarity 0.90)
• TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.90)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.89)
• SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts (AAAI 2026 · similarity 0.89)