Unlocking Zero-Shot Geospatial Reasoning via Indirect Rewards

Conference: ICML 2026
arXiv: 2510.00072
Code: https://github.com/miniHuiHui/Geo-R1
Area: Reinforcement Learning / Multimodal VLM / Geospatial Reasoning
Keywords: Indirect Reward, RLVR, Cross-View Pairing, Geospatial Reasoning, Zero-Shot Generalization

TL;DR

The authors use "whether a ground street view and a satellite image can be localized to the same coordinate" as a verifiable indirect reward, and apply two-stage post-training (CoT scaffolding + RL self-exploring) with GRPO to Qwen2.5-VL-7B. This enables the model to learn general reasoning abilities that can zero-shot transfer to 25+ geospatial tasks using only GPS metadata.

Background & Motivation

Background: Geospatial imagery (satellite/UAV/street view) is nearly unlimited, but samples with dense semantic labels are rare. Mainstream approaches (MAE/contrastive learning/RS-VLM) excel at representation and retrieval, but lack scene decomposition and reasoning abilities. Recent R1-style RL has succeeded in math/code, but the same bottleneck remains: in geospatial domains, there is no large-scale, strong direct reward signal.

Limitations of Prior Work: (1) SFT is constrained by task distribution, leading to narrow learning and OOD collapse; (2) Fully supervised detection/segmentation/VQA annotation is expensive; (3) Existing R1-style works (e.g., GLOBE) still use "geolocation" as the only direct reward, lacking a unified principle for why such tasks induce reasoning.

Key Challenge: Metadata (coordinates, timestamps) is easy to obtain and seemingly unrelated to "complex visual reasoning," but if designed properly, its verifiability can serve as the reward basis for proxy tasks.

Goal: (1) Demonstrate that "indirect verifiable rewards" can induce complex, transferable geospatial reasoning; (2) Provide theoretical explanation for when indirect rewards are effective; (3) Construct a reproducible RLVR framework and conduct large-scale OOD validation on 25+ tasks.

Key Insight: Use cross-view pairing between "street view ↔ satellite image" as a proxy task—the challenge is that it requires transferable geometric semantics such as object geometry, shadow direction, and building layout to accomplish the match.

Core Idea: Cross-View Pairing provides a binary verifiable reward. Combined with hard negatives, it blocks shortcut solutions based on nuisance features, forcing the model to learn view-invariant geometric semantics \(\Phi\), thereby acquiring general zero-shot geospatial reasoning ability.

Method

Overall Architecture

Geo-R1 uses Qwen2.5-VL-7B as the base and applies two-stage post-training: Stage 1 is Geospatial Thinking Scaffolding, using 12.6K high-quality CoT data derived from CV-Cities for SFT, injecting a unified geospatial reasoning template (observe visual cues → cross-view evidence → associate climate/geographical knowledge → draw conclusion). Stage 2 is Reasoning Elevation via Indirect Signals, switching the training objective to Cross-View Pairing: given a ground panorama \(I_g\) and \(k\) candidate satellite images \(\mathcal S=\{I_s^1,\dots,I_s^k\}\) (1 positive + \(k-1\) hard negatives from the neighborhood), the model uses CoT reasoning to select the positive. Rewards are optimized using GRPO's group-relative form for the composite reward \(r=\lambda_{\mathrm{acc}}r_{\mathrm{acc}}+\lambda_{\mathrm{fmt}}r_{\mathrm{fmt}}+\lambda_{\mathrm{len}}r_{\mathrm{len}}+\lambda_{\mathrm{rep}}r_{\mathrm{rep}}\). Both SFT and RL use full-parameter fine-tuning, 8×H100, leveraging LLama-Factory + VLM-R1 + vLLM for accelerated inference.

Key Designs

1. Cross-View Pairing + Hard-Negative Bottleneck:

   • Function: Constructs a proxy task that is verifiable, scalable, and enforces learning of view-invariant features.
   • Mechanism: Decompose image information into view-invariant geometric semantics \(\Phi(I)\) and modality-specific nuisance factors \(N(I)\). Hard negatives are sampled from the same spatiotemporal neighborhood, ensuring \(\mathcal I(Y;N(\mathcal S))=0\), so any strategy relying only on nuisance yields maximum entropy \(\mathcal H(Y|\pi_{shortcut})=\log K\) (Theorem 3.1). Geo-R1's binary reward \(r_{acc}\in\{-1,+1\}\) suffices to maximize \(\mathcal I(C;Y|\mathcal S)\Leftrightarrow\mathcal I(C;\Phi(I_s^*))\), i.e., the reasoning chain \(C\) must encode \(\Phi\).
   • Design Motivation: Formalizes the question of whether "indirect rewards can induce reasoning" as a falsifiable theorem, and provides the "difficulty gap \(\Delta\mathcal H=\log K\)" as the reasoning margin RL must overcome.

2. CoT Scaffolding: Single-Template Paradigm Injection:

   • Function: Prevents RL cold-start collapse without introducing catastrophic forgetting from SFT task diversity.
   • Mechanism: Synthesizes 12.6K reasoning traces from CV-Cities, all following the template "analyze visual cues → cross-view evidence → associate geographical knowledge → output answer"; introduces a Fact-Check Engine to verify key entities (coordinates/city names) with metadata, ensuring the scaffold does not teach factual errors.
   • Design Motivation: Traditional SFT cold start tends to cover many tasks for diversity, but this harms subsequent RL; this work takes the opposite approach—injecting only one reasoning paradigm and letting RL explore specific abilities.

3. GRPO + Composite Verifiable Reward:

   • Function: Produces structurally compliant, appropriately long, non-repetitive, and highly accurate reasoning chains under outcome-only rewards.
   • Mechanism: Uses verifiable \(r_{\mathrm{acc}}\) as the main signal, with format regularization \(r_{\mathrm{fmt}}\), length regularization \(r_{\mathrm{len}}\), and repetition penalty \(r_{\mathrm{rep}}\); GRPO is used for group-relative advantage normalization to stabilize long-horizon updates; no process reward is introduced at intermediate steps, maintaining a purely outcome-based, verifiable process.
   • Design Motivation: Pure outcome rewards are cheap (no process annotation needed) and give the model full freedom to explore; regularization prevents the model from generating abnormally long/repetitive/unformatted outputs to game the reward.

Loss & Training

Stage 1 uses standard SFT cross-entropy with Fact-Check filtering; Stage 2 uses GRPO, with RL from base yielding Geo-R1-Zero, and RL from Geo-SFT yielding Geo-R1; full-parameter fine-tuning, 8×H100; inference accelerated by vLLM; hard negatives are sampled from the same spatiotemporal neighborhood to ensure nuisance is indistinguishable.

Key Experimental Results

Main Results

Benchmark / Task	Metric	Qwen2.5-VL-7B (base)	Geo-SFT	Geo-R1-Zero	Geo-R1
In-distribution Cross-View Pairing	Accuracy	19.0%	23.1%	78.1%	82.4%
Same as above	Completion length	204.6	1127.6	587.4	378.8
GeoChain (13 tasks OOD)	Avg accuracy	baseline	—	—	Significantly higher than baseline (Fig. 1)
IMAGEO-GSS Global 6152	City / Country Acc	—	—	—	0.3272 / 0.8146 (vs GeoCLIP 0.1086 / 0.6361)
IMAGEO-GSS	Mean / Median Dist (km)	—	—	—	568.32 / 69.40 (vs GeoCLIP 943.48 / 266.90)

Ablation Study

Configuration	Key Observation	Interpretation
SFT only (Geo-SFT)	Only 4.1% higher than base, nearly random	Positive SFT cannot capture indirect signal
RL from base only (Geo-R1-Zero)	78.1%, +59% over base	Indirect reward alone can drive learning
SFT + RL (Geo-R1)	82.4%, shorter and more stable outputs	Scaffolding provides compliant reasoning template
MP16-Reason (OOD)	Nearly matches fully supervised GLOBE-7B (1km Street 17.98 vs 17.99; 2500km Continent 93.56 vs 92.52)	Indirect reward can rival direct reward
RSTeller Satellite Geolocation	Comparable to o4-mini, surpasses base	Cross-view proxy task generalizes to pure satellite view
GeoBench-VLM Satellite Understanding	Surpasses expert models like GeoChat / EarthDial	Reasoning ability transfers to fine-grained RS tasks

Key Findings

• Indirect reward training spontaneously induces high-level semantic concepts: MP16-Reason reasoning trace wordclouds show frequent concepts like "architecture," "vegetation," "climate," and "analyzing," indicating the model is truly reasoning about the physical world rather than pattern matching.
• A single proxy task unlocks broad capabilities: Training only on Cross-View Pairing enables significant zero-shot improvements on 25+ entirely different downstream tasks (VQA, geolocation, disaster assessment, land use, etc.).
• Geo-R1 also outperforms the base on satellite-view tasks, confirming that the cross-view task implicitly builds a "ground → overhead" back-projection logic, making the model a more general aerial interpreter.

Highlights & Insights

• The concept of "verifiable indirect reward" can be extended to any rare domain with "abundant raw data + sparse metadata" (e.g., medical imaging + DICOM metadata, chemical structures + reaction temperatures), representing a new RLVR paradigm.
• The hard-negative bottleneck strictly blocks shortcuts, and the entropy gap \(\Delta\mathcal H=\log K\) quantifies the reasoning margin—a rare example of theory-engineering coupling.
• Counterintuitively, "SFT should be narrow, not broad": scaffolding SFT with only one reasoning template yields the best RL warm-up and minimal forgetting.
• Even the simplest binary reward \(\{-1,+1\}\) can drive complex reasoning, showing that process reward is not essential—of great engineering value for cost-sensitive domains.

Limitations & Future Work

• The proxy task depends on "paired ground + satellite images"; finding equivalent cross-view structures in other rare domains is not straightforward;
• Only ground/aerial two views are demonstrated; not yet extended to SAR/LiDAR/multi-temporal or more complex modalities;
• Single 7B model scale; stability of outcome-only RL at 30B+ scale remains unverified;
• Still lags behind closed-source o3 on IMAGEO-GSS, attributed by the authors to differences in parameter count and RL investment.

Related Work & Insights

• vs GLOBE (Li et al. 2025a): GLOBE uses direct geolocation reward for MP16-Reason training; Geo-R1 never sees the MP16-Reason training set and achieves nearly equal performance using only cross-view pairing indirect reward, making it a strong alternative to direct reward approaches.
• vs GeoCLIP / RFM-YFCC: Traditional retrieval/representation learning excels at nearest neighbor localization but lacks reasoning and cross-task transfer; Geo-R1 comprehensively outperforms on IMAGEO-GSS.
• vs GeoReasoner / GeoChat / EarthDial: These methods rely on large-scale task-specific supervision; Geo-R1 surpasses them zero-shot with a single proxy task, inspiring the idea that "task breadth can be induced by proxy tasks rather than exhaustive supervision."
• vs DeepSeek-R1 style math/code RLVR: This work demonstrates that the RLVR paradigm can extend to verifiable but weakly related proxy rewards, providing the first clear blueprint for "how to transfer R1 to rare domains."

Rating

• Novelty: ⭐⭐⭐⭐⭐ "Indirect verifiable reward" + hard-negative bottleneck is a novel paradigm for geospatial RLVR with cross-domain methodological significance.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 25+ downstream OOD tasks + multiple benchmarks + horizontal comparison with GLOBE / GeoCLIP / o3, extremely comprehensive.
• Writing Quality: ⭐⭐⭐⭐ Clean theory-engineering integration, clear formulas and remarks; some details require appendix reference.
• Value: ⭐⭐⭐⭐⭐ Provides a reproducible RLVR blueprint for rare domains with abundant data but few labels, directly applicable to medical/climate/robotics fields.

Related Papers

• [ICLR 2026] \(\textbf{Re}^{2}\): Unlocking LLM Reasoning via Reinforcement Learning with Re-solving
• [ICML 2025] Zero-Shot Generalization of Vision-Based RL Without Data Augmentation
• [NeurIPS 2025] Reasoning Gym: Reasoning Environments for Reinforcement Learning with Verifiable Rewards
• [ICML 2025] Pessimism Principle Can Be Effective: Towards a Framework for Zero-Shot Transfer RL
• [ICML 2025] Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination