# GRASP: Geospatial pixel Reasoning viA Structured Policy learning

arXiv: https://arxiv.org/abs/2508.17102 · HTML: https://mlsanity.com/paper/2508.17102/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- Placed in Spatial and embodied reasoning, one of 17 directions in LLM reasoning.
- Builds on 17 of its 54 references that are in the map.
- First posted on arXiv 23 Aug 2025.

## Placed in

- LLM reasoning / [Spatial and embodied reasoning](https://mlsanity.com/lr/spatial-and-embodied-reasoning/)
- Multimodal LLMs / vision-language models / [Embodied and 3D vision-language models](https://mlsanity.com/mllm/embodied-and-3d-vision-language-models/)

## Builds on (most cited first, 17 of 17)

- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Qwen2.5-VL Technical Report](https://mlsanity.com/paper/2502.13923/) (2025)
- [Qwen Technical Report](https://mlsanity.com/paper/2309.16609/) (2023)
- [SAM 2: Segment Anything in Images and Videos](https://mlsanity.com/paper/2408.00714/) (2024)
- [HybridFlow: A Flexible and Efficient RLHF Framework](https://mlsanity.com/paper/2409.19256/) (2024)
- [Gemini: A Family of Highly Capable Multimodal Models](https://mlsanity.com/paper/2312.11805/) (2023)
- [DeepSeek-VL: Towards Real-World Vision-Language Understanding](https://mlsanity.com/paper/2403.05525/) (2024)
- [LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation](https://mlsanity.com/paper/2110.08733/) (2021)
- [Seg-Zero: Reasoning-Chain Guided Segmentation via Cognitive Reinforcement](https://mlsanity.com/paper/2503.06520/) (2025)
- [LISA++: An Improved Baseline for Reasoning Segmentation with Large Language Model](https://mlsanity.com/paper/2312.17240/) (2023)
- [GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing](https://mlsanity.com/paper/2501.13925/) (2025)
- [GeoGround: A Unified Large Vision-Language Model for Remote Sensing Visual Grounding](https://mlsanity.com/paper/2411.11904/) (2024)
- [SegEarth-R1: Geospatial Pixel Reasoning via Large Language Model](https://mlsanity.com/paper/2504.09644/) (2025)
- [RSRefSeg: Referring Remote Sensing Image Segmentation with Foundation Models](https://mlsanity.com/paper/2501.06809/) (2025)
- [RSRefSeg 2: Decoupling Referring Remote Sensing Image Segmentation with Foundation Models](https://mlsanity.com/paper/2507.06231/) (2025)
- [GeoPix: Multi-Modal Large Language Model for Pixel-level Image Understanding in Remote Sensing](https://mlsanity.com/paper/2501.06828/) (2025)

## Built on by (most cited first, 4 of 4)

- [Look Where It Matters: Training-Free Ultra-HR Remote Sensing VQA via Adaptive Zoom Search](https://mlsanity.com/paper/2511.20460/) (2025)
- [RSGround-R1: Rethinking Remote Sensing Visual Grounding through Spatial Reasoning](https://mlsanity.com/paper/2601.21634/) (2026)
- [WeaveEarth: Structured Evidence Construction and Reasoning for Training-Free UHR Remote Sensing Understanding](https://mlsanity.com/paper/2607.10120/) (2026)
- [AgroVG: A Large-Scale Multi-Source Benchmark for Agricultural Visual Grounding](https://mlsanity.com/paper/2605.22034/) (2026)

## Abstract (arXiv)

Geospatial pixel reasoning aims to generate segmentation masks in remote sensing imagery directly from natural-language instructions. Most existing approaches follow a paradigm that fine-tunes multimodal large language models under supervision with dense pixel-level masks as ground truth. While effective within the training data distribution, this design suffers from two main drawbacks: (1) the high cost of large-scale dense mask annotation, and (2) the limited generalization capability of supervised fine-tuning in out-of-domain scenarios. To address these issues, we propose GRASP, a structured policy-learning framework that integrates a multimodal large language model with a pretrained segmentation model in a cascaded manner. To enhance generalization, we introduce PRIME, a training paradigm that replaces supervised fine-tuning with reinforcement learning to better align reasoning and grounding behaviors with task objectives. To reduce annotation costs, we design BoP-Rewards, which substitutes dense mask labels with bounding box and positive points. It further verifies outputs through two complementary signals: format, which constrains the reasoning and grounding structure to remain syntactically parsable, and accuracy, which evaluates the quality of predicted boxes and points. For evaluation, we train our method and all baselines on EarthReason and GeoPixInstruct, constructing an in-domain benchmark by merging their test sets. We further release GRASP-1k, a fully out-of-domain benchmark with reasoning-intensive queries, reasoning traces, and fine-grained masks. Experimental results demonstrate state-of-the-art (SOTA) in-domain performance and up to 54% improvement in out-of-domain scenarios, confirming that reinforcement learning with cost-aware rewards provides a robust and scalable paradigm for geospatial pixel reasoning. All code and datasets will be released publicly.