Source-linked AI summary
RemoteZero: Geospatial Reasoning with Zero Labels
Liang Yao, Fan Liu, Shengxiang Xu, Chuanyi Zhang, Rui Min, Shimin Di, Yuhui Zheng
TL;DR
Geospatial reasoning lacks scalable label-free rewards because reinforcement methods still rely on human-provided solution labels. RemoteZero crops predicted regions for semantic verification, uses that signal for GRPO and self-evolution, and reaches 71.29% Acc@0.5 on EarthReason, 3.18 points above the strongest supervised baseline.
Problem
Reinforcement-based geospatial reasoning still requires human-provided target labels for rewards, limiting use on large-scale unlabeled Earth observation data.
Method
RemoteZero replaces annotation-dependent rewards with semantic verification of padded predicted-region crops and reuses previous-round models as verifiers for self-evolution.
Results
71.29% Acc@0.5 on EarthReason surpasses the strongest supervised baseline by 3.18 percentage points without labels for policy optimization.
Takeaways & Limitations
RemoteZero provides a label-free route for reinforcement-based geospatial reasoning and supports continual improvement from unlabeled data.
Takeaways & Limitations
The verifier interface is agnostic to whether verification comes from an external superior model or the policy itself.
Abstract
from arXiv · showhide
Geospatial reasoning requires models to identify image regions that satisfy complex and often implicit user intents. Recent reinforcement learning approaches improve reasoning without manually annotated reasoning traces, but still require human-provided target labels to construct rewards, limiting their use on large-scale unlabeled Earth observation data. We introduce RemoteZero, a label-free framework for reinforcement-based geospatial reasoning. Our key observation is twofold: MLLMs are often more reliable at evaluating candidates than generating solutions, while aerial imagery reduces interference in region-level verification. Therefore, RemoteZero converts each predicted region into a visual crop and uses its semantic consistency with the query as an intrinsic reward for GRPO optimization. This formulation removes the need for human-provided solution labels and further supports iterative self-evolution by reusing previous-round models as verifiers. Experiments show that RemoteZero outperforms strong supervised baselines, applies effectively to other Earth observation tasks, and continues to improve through self-evolution as the training data expand. We hope this direction can broadly benefit the Earth observation community.
1. Introduction
Geospatial reasoning targets implicit functional intents rather than predefined objects, while existing reinforcement approaches still depend on human-provided solution labels. RemoteZero removes this dependency through semantic verification, self-evolution, and label-free optimization on Earth observation data.
- Geospatial reasoning requires identifying regions that satisfy implicit functional needs, jointly interpreting intended use, spatial constraints, and scene evidence.
- Existing approaches use manually curated image–reasoning–mask triplets or reinforcement learning with target labels, limiting scalability and potentially encouraging fixed reasoning patterns.
- RemoteZero crops each predicted region and evaluates its semantic consistency with the query to provide an intrinsic GRPO reward without human-labeled coordinates.
- RemoteZero supports verifier-based distillation and iterative self-evolution by reusing previous-round models as verifiers.
- 71.29% test Acc@0.5 on EarthReason surpasses the strongest supervised baseline by 3.18 percentage points without labels for policy optimization.
- The method transfers to visual grounding and referring segmentation and benefits from increasing unlabeled training data across self-evolution rounds.
2. Motivation
RemoteZero is motivated by an asymmetry in which MLLMs can verify specified regions more reliably than they can generate them. Overhead imagery makes candidate crops easier to interpret, enabling semantic verification as a label-free training signal.
- Generating a valid structured region requires open-ended search over outputs, whereas verification evaluates a specified candidate through a scalar or binary judgment.
- RemoteZero combines a padded candidate crop, verifier score, and area penalty into an intrinsic GRPO reward without ground-truth coordinates.
- On EarthReason, Qwen3-VL-8B achieves 43.88% Acc@0.5 when generating regions directly but 66.73% verification accuracy when judging ground-truth regions.
- Overhead or near-nadir imagery reduces depth ambiguity, severe occlusion, and cross-object overlap, supporting more reliable region-level verification.
- Moderate contextual padding preserves nearby roads, buildings, fields, or open space while keeping semantic evidence attributable to the proposed region.
3. RemoteZero
RemoteZero replaces annotation-dependent spatial rewards with crop-based semantic verification, enabling label-free GRPO training and iterative self-evolution. Its verifier can be external initially or reused from preceding policy rounds.
- Problem formulation: RemoteZero replaces human-annotated spatial matching rewards with semantic verification of each predicted region against the query.The reformulated reward enables policy learning without spatial labels.
- Policy optimization: The policy generates a reasoning chain and bounding box, guided by subsequent reward rather than ground-truth coordinates.GRPO fine-tunes the policy from sampled outputs and their rewards.
- Generate-Crop-Verify: A deterministic crop with relaxed contextual padding isolates the predicted region while retaining nearby roads, terrain, or other verification cues.The crop is passed to a verifier that outputs confidence for semantic entailment between the region and query.
- Reward design: The final GRPO reward combines verifier confidence with an area penalty to discourage overly large crops that sacrifice spatial precision.The penalty uses λ for strength and τ for the acceptable area threshold.
- Self-evolution: Self-evolution reuses each completed policy as the frozen verifier for the next training round, so external verification is needed only for initialization.Subsequent rounds use the same crop-based verification and area regularization without coordinate annotations.
4.1. Experimental Setup
Experiments evaluate RemoteZero on EarthReason and two additional spatial prediction tasks, using several MLLM and specialized supervised baselines.
- Implementation: The implementation uses Qwen3-VL-8B-Instruct with GRPO and LoRA fine-tuning.
- Datasets: EarthReason evaluates geospatial reasoning, while DIOR-RSVG and RRSIS-D assess visual grounding and referring segmentation.
- Baselines: The EarthReason comparison includes general-purpose MLLMs, remote-sensing MLLMs, and supervised geospatial reasoning methods.
- Baselines: Visual grounding baselines include SkyEyeGPT, GeoChat, LHRS-Bot, RemoteSAM, EarthDial, EarthGPT, and GeoZero.
- Baselines: Referring segmentation is compared against LISA, PixelLM, Text4Seg++, GeoGround, SegEarth-R1, and ProVG.
4.2. Main Results
RemoteZero substantially improves implicit geospatial localization over general-purpose and supervised baselines, while its verifier remains less effective at calibrating precise spatial extents.
- EarthReason results: 48.10% test Acc@0.5 is achieved by the strongest general-purpose baseline, Qwen3-VL-8B, while VLM-R1 reaches 33.31%.
- EarthReason results: 71.29% test Acc@0.5 is achieved after self-evolution, surpassing the strongest supervised baseline by 3.18 percentage points.
- EarthReason results: RemoteZero* reaches 65.05% test Acc@0.5 without human-provided solution labels for policy optimization, approaching supervised RemoteReasoner.
- EarthReason results: RemoteZero obtains lower test gIoU than RemoteReasoner, indicating weaker calibration of precise spatial extents despite semantically correct localization.
4.3. Ablation Studies
Ablations show that reward design, contextual cropping, and verifier selection materially affect training, with moderate context and area-aware rewards improving localization.
- Reward functions: 69.96% validation and 71.29% test Acc@0.5 result when an area penalty is added to semantic verification.
- Reward functions: Semantic verification alone reaches 65.20% validation and 65.88% test Acc@0.5 but can favor overly large boxes.
- Cropping strategy: 15% contextual padding achieves the best cropping results, reaching 69.96% validation and 71.29% test Acc@0.5.
- Cropping strategy: Strict cropping reaches 64.61% validation and 65.13% test Acc@0.5, while 20% padding lowers accuracy by introducing irrelevant regions.
- Verifier models: Qwen2.5-VL-72B yields the best downstream validation and test results, whereas Qwen3-VL-32B obtains the highest verification score.
- Verifier models: Qwen3-VL-8B is adopted as the default verifier because it offers a favorable trade-off between verification quality and training efficiency.
4.4. Further Analysis
Further analyses examine reward sensitivity, transfer to other spatial tasks, and iterative improvement, showing benefits from moderate context, self-evolution, and expanding unlabeled data.
- Verification behavior: The verifier assigns lower rewards to shifted boxes and target-free regions than to ground-truth regions, showing sensitivity to spatial relevance.
- Transfer to other tasks: RemoteZero achieves 76.0 Pr@0.5 and 68.9 mIoU on visual grounding, outperforming most specialized grounding methods.
- Transfer to other tasks: Bounding boxes from the label-free grounding model are used as prompts for SAM 2 in referring segmentation without segmentation labels.
- Self-evolution: 8.26 and 8.10 percentage-point gains in validation and test Acc@0.5 occur from the first to the sixth self-evolution round.
- Self-evolution: The sixth round reaches 69.96% validation and 71.29% test Acc@0.5, while gains from the fifth to sixth round shrink to 0.18 and 0.52 points.
- Data scaling: With continually expanding training data, verifier and policy performance continue improving beyond the plateau observed under a fixed training set.
- Qualitative case study: A successful case receives verifier reward 0.988, while a semantically mismatched oversized beach prediction receives zero reward.
5. Related Work
Remote sensing MLLMs have progressed from domain-specific captioning and grounded dialogue toward unified interpretation, fine-grained interaction, and complex geospatial reasoning. Recent work increasingly uses reinforcement learning and verifiable reasoning for implicit queries and diverse observation tasks.
- Remote sensing MLLMs evolved from domain-specific captioning and grounded dialogue toward unified multi-sensor interpretation and instruction tuning.
- Later systems improved interaction and granularity through relational reasoning, VGI-enhanced training, fine-grained instruction tuning, and visual prompting.
- Geospatial reasoning work moved beyond standard perception to address implicit queries through pixel-level reasoning and unified reinforcement-learning workflows.
- Recent methods increasingly combine reinforcement fine-tuning or chain-of-thought reasoning with task-aware rewards and perceptually grounded, verifiable outputs.
6. Conclusion
RemoteZero is a label-free framework for reinforcement-based geospatial reasoning that replaces human-provided solution labels with intrinsic semantic verification. On EarthReason, it reaches 71.29% Acc@0.5, surpasses the strongest supervised baseline by 3.18 percentage points, and supports self-evolution through reused verifiers.
- RemoteZero provides a label-free framework for reinforcement-based geospatial reasoning.
- RemoteZero replaces human-provided solution labels with intrinsic semantic verification and reuses previous-round models as verifiers for self-evolution.
- 71.29% Acc@0.5 on EarthReason surpasses the strongest supervised baseline by 3.18 percentage points.