GeoZero: Zero-Supervision Geospatial Reasoning

• GeoZero is a zero-supervision geospatial reasoning framework that eliminates chain-of-thought supervision through answer-anchored reinforcement learning.
• It integrates a large instruction-tuning corpus with a hard-sample RL phase to trigger explicit reasoning and achieve high remote-sensing task performance.
• The framework’s dual approach—answer-anchored and verification-anchored—demonstrates how intrinsic rewards can effectively replace curated spatial annotations.

Searching arXiv for GeoZero and closely related papers to ground the article. GeoZero denotes a class of training paradigms for geospatial reasoning under minimal or zero human supervision, and more specifically a 2025 framework for remote-sensing multimodal LLMs (MLLMs) that eliminates predefined chain-of-thought supervision while incentivizing reasoning through answer-based reinforcement learning (Wang et al., 27 Nov 2025). In the narrow sense established by the paper "GeoZero: Incentivizing Reasoning from Scratch on Geospatial Scenes" (Wang et al., 27 Nov 2025), GeoZero combines a large instruction-tuning corpus, a hard-sample reinforcement-learning set, and Answer-Anchored Group Relative Policy Optimization (A$^2$GRPO) to induce explicit geospatial reasoning without any human-written or LLM-written CoT traces. In a broader sense, the term also names a "GeoZero" paradigm instantiated by RemoteZero, where geospatial localization is trained without human box or coordinate annotations, replacing geometric supervision with intrinsic semantic verification (Yao et al., 6 May 2026). Taken together, these works define GeoZero as an annotation-minimal regime for geospatial reasoning in which supervision is anchored in answers or intrinsic verification rather than curated reasoning traces or spatial labels.

1. Conceptual Scope and Historical Placement

GeoZero arose in response to a specific limitation in remote-sensing MLLMs: strong supervised fine-tuning improves instruction following and task accuracy, but tends to favor direct answer patterning over explicit reasoning, especially on complex spatial tasks (Wang et al., 27 Nov 2025). The GeoZero framework addresses this by asking whether an MLLM can acquire geospatial reasoning capability without any predefined CoT supervision and by answering that question through a two-stage SFT-plus-RL pipeline (Wang et al., 27 Nov 2025).

The central contrast is with "cold-start" reasoning pipelines that first synthesize or curate CoT traces and then supervise models to imitate them. GeoZero rejects this requirement. It uses supervised fine-tuning only to impart geospatial knowledge and task format regularity, and then relies on reinforcement learning over hard geospatial examples to make reasoning emerge (Wang et al., 27 Nov 2025). This places GeoZero adjacent to general self-evolving and verifier-free learning frameworks, but specialized to remote sensing and multimodal geospatial tasks.

A related but distinct development is RemoteZero, which the paper explicitly describes as a concrete instantiation of a "GeoZero" paradigm for localization: geospatial reasoning and localization trained without any human box or coordinate annotations (Yao et al., 6 May 2026). This suggests that "GeoZero" is best understood not only as the name of a single framework, but also as a family of geospatial learning regimes characterized by the removal of hand-authored intermediate supervision.

2. GeoZero Framework for Reasoning Without CoT Supervision

In the 2025 formulation, GeoZero consists of three components: GeoZero-Instruct, GeoZero-Hard, and A$^2$GRPO (Wang et al., 27 Nov 2025). The base model is Qwen3-VL-8B-Instruct, with its vision encoder frozen and the remaining trainable components adapted through LoRA (Wang et al., 27 Nov 2025).

The supervised phase uses GeoZero-Instruct, approximately $610\text{k}$ samples, to teach canonical geospatial tasks without any CoT traces (Wang et al., 27 Nov 2025). The reinforcement-learning phase uses GeoZero-Hard, approximately $20\text{k}$ hard examples, to force deeper reasoning on difficult cases across the same task families (Wang et al., 27 Nov 2025). The hard set is image-disjoint from the SFT set, so the RL stage does not simply revisit previously memorized scenes (Wang et al., 27 Nov 2025).

The resulting training logic is asymmetric. Supervised fine-tuning provides remote-sensing vocabulary, task conventions, and baseline competence, but does not teach explicit reasoning structure. Reinforcement learning then supplies the incentive to produce "reasoning then answer" behavior under a reward that values both answer quality and reasoning quality, while never comparing model outputs to reference CoTs (Wang et al., 27 Nov 2025).

This design differs from generic self-play systems such as G-Zero, which targets open-ended language generation through a Generator–Proposer co-evolutionary loop and a Hint-$\delta$ intrinsic reward (Huang et al., 11 May 2026). GeoZero does not use proposer-generated tasks or hint-conditioned preference pairs. Instead, it keeps the remote-sensing task structure fixed and shapes behavior through task-specific correctness rewards and answer-anchored policy regularization (Wang et al., 27 Nov 2025).

3. Data Construction: GeoZero-Instruct and GeoZero-Hard

GeoZero begins from GeoZero-Raw, approximately $754{,}749$ instances aggregated from remote-sensing datasets spanning scene classification, visual grounding, visual question answering, and image captioning (Wang et al., 27 Nov 2025). All samples are converted into instruction-following dialogues with task tags such as [cls], [grounding], [vqa], and [caption], and task-specific textual hints are included with $50\%$ probability using more than 20 prompt variants per task (Wang et al., 27 Nov 2025).

The four principal task categories are as follows.

Task type	Role in GeoZero-Raw	Example source datasets
Scene Classification (SC)	aerial scene label prediction	AID, RESISC-45, EuroSAT, UCM
Visual Grounding (VG)	phrase-to-box localization	RSVG, DIOR-RSVG, VRSBench VG
Visual Question Answering (VQA)	remote-sensing QA	RSVQA-HR, RSVQA-LR, VRSBench
Image Captioning (IC)	scene description	SkyEye968k, VRSBench caption splits

GeoZero-Hard is built by automatic hard-sample mining with a Data Filtering Model (DFM) based on Qwen2.5-VL-7B-Instruct fine-tuned on GeoZero-Raw for 3 epochs with a frozen vision encoder and LoRA (Wang et al., 27 Nov 2025). For each sample, the DFM generates predictions and task-specific correctness is computed using exact or substring text matching for SC and VQA, $\mathrm{IoU} \ge 0.5$ for VG, and word-set $F_1 \ge 0.6$ for IC (Wang et al., 27 Nov 2025). Incorrect cases are retained as candidates. In a second stage, three stochastic generations are sampled per candidate, the mean correctness $Acc$ is computed, and difficulty is defined as $^2$0; top-ranked examples are then selected with task balance to form GeoZero-Hard (Wang et al., 27 Nov 2025).

GeoZero-Instruct is obtained by removing from GeoZero-Raw any sample whose image appears in GeoZero-Hard, leaving approximately $^2$1 SFT examples (Wang et al., 27 Nov 2025). This image-level separation is technically important because it prevents RL from exploiting overlap between "knowledge acquisition" and "reasoning pressure" stages.

4. A$^2$2GRPO and Reward Design

GeoZero’s reinforcement-learning stage uses Answer-Anchored Group Relative Policy Optimization (A$^2$3GRPO), a modification of GRPO in which rewards are computed from answer quality and reasoning quality, while KL-style regularization is applied only to answer tokens (Wang et al., 27 Nov 2025). The model is prompted to reason first and then emit a final answer inside <answer> ... </answer>; a binary mask $^2$4 marks answer tokens (Wang et al., 27 Nov 2025).

The total reward is

$^2$5

where $^2$6 is the task-specific answer reward and $^2$7 is the Answer-Modulated Thinking Reward (AMTR) (Wang et al., 27 Nov 2025).

For answer quality, GeoZero uses different metrics by task (Wang et al., 27 Nov 2025):

• SC and VQA: semantic similarity from bge-small-en-v1.5 embeddings,

$^2$8

• VG: bounding-box IoU,

$^2$9

• IC: a weighted combination

$610\text{k}$0

with $610\text{k}$1, $610\text{k}$2, $610\text{k}$3, and $610\text{k}$4 (Wang et al., 27 Nov 2025).

The reasoning reward is gated by answer quality: $610\text{k}$5 with $610\text{k}$6 and $610\text{k}$7 (Wang et al., 27 Nov 2025). This means poor answers suppress the thinking reward, so verbose but incorrect reasoning is not advantageous.

The thinking-quality score $610\text{k}$8 is defined by

$610\text{k}$9

where $20\text{k}$0 combines a normalized length score $20\text{k}$1, a redundancy penalty $20\text{k}$2, and an answer-overlap penalty $20\text{k}$3, while $20\text{k}$4 is a semantic diversity bonus based on sentence-embedding cosine differences (Wang et al., 27 Nov 2025). The length thresholds are $20\text{k}$5, $20\text{k}$6, $20\text{k}$7, and $20\text{k}$8; the redundancy thresholds are $20\text{k}$9 and $\delta$0; and the answer-overlap threshold is $\delta$1 (Wang et al., 27 Nov 2025).

The A$\delta$2GRPO objective is

$\delta$3

where the KL-like regularization term is multiplied by $\delta$4, so it applies only to answer tokens (Wang et al., 27 Nov 2025).

This selective regularization is the formal core of the "answer-anchored" design. The answer remains close to the SFT reference distribution, while the reasoning segment is free to diverge and explore. A plausible implication is that GeoZero treats reasoning not as a supervised latent program but as an unconstrained behavior shaped only by downstream utility.

5. Emergent Reasoning and Empirical Behavior

GeoZero reports that explicit reasoning does not emerge reliably from prompting alone and is often suppressed by conventional SFT (Wang et al., 27 Nov 2025). In the supplement, Qwen3-VL-8B-Instruct prompted to reason shows some reasoning on UCM but almost none on RSVG, and after SFT on GeoZero-Instruct the "thinking activation rate" becomes $\delta$5 (Wang et al., 27 Nov 2025). GeoZero’s RL stage reverses this behavior.

A key ablation compares training paradigms on UCM and RSVG (Wang et al., 27 Nov 2025). The pattern is consistent: SFT-only produces strong task accuracy but no reasoning; RL-only can activate reasoning but performs poorly; SFT + RL on GeoZero-Hard activates reasoning while retaining strong performance; and GeoZero + RFT yields both explicit reasoning and best accuracy (Wang et al., 27 Nov 2025).

The paper also shows that random RL data are insufficient. Replacing GeoZero-Hard with random samples from GeoZero-Raw results in little or no reasoning emergence before reinforcement fine-tuning, whereas hard-sample RL triggers reasoning directly and improves visual grounding markedly (Wang et al., 27 Nov 2025). This indicates that the difficulty distribution of RL data is not incidental but constitutive of the method.

Ablations of A$\delta$6GRPO further show that both the thinking reward and the token mask are required (Wang et al., 27 Nov 2025). Without AMTR, the model tends not to reason. Without masking, reasoning exploration is constrained by the reference model. Full A$\delta$7GRPO is the only setting that consistently yields both reasoning and strong visual grounding accuracy (Wang et al., 27 Nov 2025).

On DIOR-RSVG, GeoZero analyzes the relation between reasoning quality and task success. Accuracy increases with reasoning length up to roughly 40–80 words and correlates positively with the structural score $\delta$8 and the total thinking-quality score $\delta$9, while very long or semantically diffuse reasoning can reduce performance (Wang et al., 27 Nov 2025). This suggests that GeoZero’s reward is not merely inducing format compliance; it is shaping a specific regime of effective reasoning behavior.

6. Performance Across Remote-Sensing Benchmarks

GeoZero is evaluated on scene classification, visual grounding, VQA, image captioning, and two broader geospatial reasoning benchmarks (Wang et al., 27 Nov 2025). The following condensed summary reflects the reported metrics.

Benchmark family	GeoZero result	Notes
Scene Classification	UCM 95.48%, AID 97.30%, WHU-RS19 98.06%	with RFT (Wang et al., 27 Nov 2025)
Visual Grounding	RSVG 50.04%, DIOR-RSVG 79.43%, VRS-VG 73.32%	with RFT (Wang et al., 27 Nov 2025)
VQA	RSVQA-HR Test1 Presence 91.16%, Compare 91.01%	with RFT (Wang et al., 27 Nov 2025)
Image Captioning	strong CIDEr results, e.g. UCM-Captions 393.57	with RFT (Wang et al., 27 Nov 2025)
XLRSBench / CHOICE	48.10 / 72.06 average L3 accuracy	with RFT (Wang et al., 27 Nov 2025)

These results are reported as exceeding or matching prior specialized geospatial MLLMs and, on XLRSBench and CHOICE, exceeding GPT-4o on the cited domain benchmarks (Wang et al., 27 Nov 2025). Because the paper emphasizes that reasoning is acquired without CoT supervision, these results are presented not only as benchmark gains but as evidence that answer-anchored RL can induce broadly useful geospatial reasoning skills.

A related benchmark result from RemoteZero is also germane to the GeoZero paradigm. On EarthReason, RemoteZero (Self-Evolution) reaches Val [email protected] 69.96, Test 71.29, Val gIoU 61.54, Test 61.70, while RemoteReasoner, a supervised baseline trained with ground-truth boxes and IoU reward, reaches Val [email protected] 66.51, Test 68.11, Val gIoU 67.04, Test 69.29 (Yao et al., 6 May 2026). The implication is precise: annotation-free semantic verification can surpass box-supervised training on region correctness while lagging on boundary calibration.

7. GeoZero as a Broader Annotation-Free Geospatial Paradigm

RemoteZero expands the meaning of GeoZero from "reasoning without CoT" to "geospatial reasoning and localization trained without any human box / coordinate annotations" (Yao et al., 6 May 2026). It addresses the task of mapping a remote-sensing image $754{,}749$0 and a natural-language query $754{,}749$1 to a reasoning sequence $754{,}749$2 and a bounding box

$754{,}749$3

while assuming no ground-truth boxes during training (Yao et al., 6 May 2026).

Its central principle is the "Eye > Hand" asymmetry: MLLMs are better at verifying whether a cropped region matches a textual query than at directly generating precise coordinates (Yao et al., 6 May 2026). RemoteZero therefore replaces external geometric supervision

$754{,}749$4

with an intrinsic semantic reward

$754{,}749$5

where $754{,}749$6 is a crop operator and $754{,}749$7 is an MLLM verifier (Yao et al., 6 May 2026). The complete reward adds an area penalty,

$754{,}749$8

to prevent trivial large-box solutions (Yao et al., 6 May 2026).

Training proceeds via Group Relative Policy Optimization (GRPO) using groups of $754{,}749$9 trajectories per prompt, and crucially the objective contains no $50\%$0 or IoU-to-ground-truth term (Yao et al., 6 May 2026). RemoteZero also supports iterative self-evolution: at iteration $50\%$1, the frozen previous policy acts as verifier,

$50\%$2

so the model improves on unlabeled Earth observation data using its own verification signal (Yao et al., 6 May 2026).

This suggests a general taxonomy within GeoZero-style geospatial learning:

• Answer-anchored GeoZero: supervision only on final answers, with reasoning induced through RL and reward shaping (Wang et al., 27 Nov 2025).
• Verification-anchored GeoZero: supervision only through intrinsic semantic checking of predicted regions, without geometric labels (Yao et al., 6 May 2026).

The two share a common philosophical move: replace human-authored intermediate supervision with internally generated or task-anchored evaluative signals.

8. Relation to Other Zero-Supervision and Self-Evolution Frameworks

GeoZero is related to, but technically distinct from, several "zero" frameworks beyond remote sensing. G-Zero, for example, is a verifier-free self-play system for open-ended language generation that uses a Generator, a Proposer, and a Hint-$50\%$3 intrinsic reward derived from predictive shifts between unassisted and hint-conditioned responses (Huang et al., 11 May 2026). Its objective is broad autonomous self-improvement across unverifiable language tasks, not geospatial multimodal reasoning. Nonetheless, both G-Zero and GeoZero eliminate external judges and derive supervision from internal dynamics or answer-conditioned rewards (Huang et al., 11 May 2026, Wang et al., 27 Nov 2025).

GeoZero is also adjacent to GRPO-based optimization work in general. RemoteZero uses GRPO for localization without box supervision (Yao et al., 6 May 2026), while GeoZero adapts GRPO into A$50\%$4GRPO by selectively masking KL regularization over answer tokens (Wang et al., 27 Nov 2025). This suggests that GRPO serves as a common optimization substrate for recent "zero-supervision" paradigms, but the source of reward differs substantially across domains.

A possible misconception is that GeoZero is simply a geospatial rebranding of generic self-play or RLHF-free training. The available evidence does not support that. GeoZero depends on domain-specific data construction, task-specific reward functions, and remote-sensing benchmarks, and its central innovations are tightly bound to geospatial scene understanding rather than open-ended language generation (Wang et al., 27 Nov 2025). Conversely, RemoteZero shows that the same label can extend beyond reasoning traces to spatial endpoint supervision, but still within remote sensing (Yao et al., 6 May 2026).

9. Limitations and Open Questions

GeoZero’s limitations follow directly from the cited papers. In the 2025 reasoning framework, coverage is restricted to four major task types—SC, VG, VQA, and IC—and does not directly include change detection, multi-temporal forecasting, segmentation, or hyperspectral reasoning (Wang et al., 27 Nov 2025). The backbone is fixed to Qwen3-VL-8B-Instruct, so backbone sensitivity remains open (Wang et al., 27 Nov 2025). Reinforcement learning is computationally costly, and the definition of "hard" examples depends on a DFM whose own failure modes may bias the RL curriculum (Wang et al., 27 Nov 2025). The paper also notes that high-quality reasoning correlates with accuracy, but this does not prove every trace is faithfully causal rather than post hoc (Wang et al., 27 Nov 2025).

RemoteZero identifies a different set of limitations for GeoZero-style localization. Its reward emphasizes semantic correctness over boundary precision, which explains why [email protected] can exceed supervised baselines while gIoU remains lower (Yao et al., 6 May 2026). Self-evolution can accumulate verifier bias; crop-based verification may miss global spatial relations; GRPO with multiple generations and MLLM verification is computationally heavy; and robustness beyond EarthReason, including across SAR, multi-spectral, and cross-geographic domains, remains unvalidated (Yao et al., 6 May 2026).

These limitations point toward a broader research agenda. One direction is to combine GeoZero’s answer-anchored reasoning incentives with RemoteZero’s verification-driven spatial learning, yielding joint reasoning-and-localization agents trained with minimal annotation. Another is to introduce geometry-aware auxiliary rewards or global–local verification mechanisms to narrow the gap between semantic correctness and precise localization (Yao et al., 6 May 2026). A third is to extend GeoZero-style reasoning emergence to multi-temporal and multi-sensor settings, where "hardness" may need to be defined over temporal dynamics rather than static scenes (Wang et al., 27 Nov 2025).

In that sense, GeoZero is best viewed not as a completed doctrine but as an organizing principle for remote-sensing MLLMs: geospatial reasoning should be induced from task difficulty, answer quality, and intrinsic verification signals, rather than from expensive curated traces or dense spatial annotations (Wang et al., 27 Nov 2025, Yao et al., 6 May 2026).