---
title: 'GeoZero: Zero-Supervision Geospatial Reasoning'
url: https://www.emergentmind.com/topics/geozero
type: topic
---

# GeoZero: Zero-Supervision Geospatial Reasoning

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 large language models (MLLMs) that eliminates predefined chain-of-thought supervision while incentivizing reasoning through answer-based reinforcement learning [2511.22645]. In the narrow sense established by the paper "GeoZero: Incentivizing Reasoning from Scratch on Geospatial Scenes" [2511.22645], 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 [2605.04451]. 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 [2511.22645]. 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 [2511.22645].

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 [2511.22645]. 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 [2605.04451]. 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** [2511.22645]. The base model is **Qwen3-VL-8B-Instruct**, with its vision encoder frozen and the remaining trainable components adapted through LoRA [2511.22645].

The supervised phase uses GeoZero-Instruct, approximately \(610\text{k}\) samples, to teach canonical geospatial tasks without any CoT traces [2511.22645]. 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 [2511.22645]. The hard set is image-disjoint from the SFT set, so the RL stage does not simply revisit previously memorized scenes [2511.22645].

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 [2511.22645].

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 [2605.09959]. 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 [2511.22645].

## 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 [2511.22645]. 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 [2511.22645].

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 [2511.22645]. 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 [2511.22645]. 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 \(1-Acc\); top-ranked examples are then selected with task balance to form GeoZero-Hard [2511.22645].

GeoZero-Instruct is obtained by removing from GeoZero-Raw any sample whose image appears in GeoZero-Hard, leaving approximately \(610\text{k}\) SFT examples [2511.22645]. This image-level separation is technically important because it prevents RL from exploiting overlap between "knowledge acquisition" and "reasoning pressure" stages.

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

GeoZero’s reinforcement-learning stage uses **Answer-Anchored Group Relative Policy Optimization (A$^2$GRPO)**, 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 [2511.22645]. The model is prompted to reason first and then emit a final answer inside `<answer> ... </answer>`; a binary mask \(M_i = \{m_{i,t}\}\) marks answer tokens [2511.22645].

The total reward is
\[
r = r_a + \lambda r_t,\qquad \lambda = 0.3,
\]
where \(r_a\) is the task-specific answer reward and \(r_t\) is the **Answer-Modulated Thinking Reward (AMTR)** [2511.22645].

For answer quality, GeoZero uses different metrics by task [2511.22645]:

- **SC and VQA**: semantic similarity from **bge-small-en-v1.5** embeddings,
  \[
  r_a^{(\text{SC/VQA})} = \frac{\cos(f(x), f(y)) + 1}{2}.
  \]
- **VG**: bounding-box IoU,
  \[
  r_a^{(\text{VG})} = \mathrm{IoU}(\hat{B}, B).
  \]
- **IC**: a weighted combination
  \[
  r_a^{(\text{IC})} = w_b s_b + w_r s_r + w_m s_m + w_e s_e,
  \]
  with \(w_b=0.15\), \(w_r=0.2\), \(w_m=0.3\), and \(w_e=0.35\) [2511.22645].

The reasoning reward is gated by answer quality:
\[
r_t = \mathcal{G}(r_a)\cdot r_a \cdot s_t,
\qquad
\mathcal{G}(r_a)=\frac{1}{1+e^{-k(r_a-\tau)}},
\]
with \(k=12\) and \(\tau=0.5\) [2511.22645]. This means poor answers suppress the thinking reward, so verbose but incorrect reasoning is not advantageous.

The thinking-quality score \(s_t\) is defined by
\[
s_t = (1-w_d)\,q_t + w_d(q_t\cdot b_d), \qquad w_d=0.3,
\]
where \(q_t = l_s \cdot p_r \cdot p_a\) combines a normalized length score \(l_s\), a redundancy penalty \(p_r\), and an answer-overlap penalty \(p_a\), while \(b_d\) is a semantic diversity bonus based on sentence-embedding cosine differences [2511.22645]. The length thresholds are \(\tau_{\min}=20\), \(\tau_{\text{lo}}=40\), \(\tau_{\text{hi}}=80\), and \(\tau_{\max}=160\); the redundancy thresholds are \(\delta_r=0.15\) and \(\gamma_r=0.5\); and the answer-overlap threshold is \(\tau_a=0.3\) [2511.22645].

The A$^2$GRPO objective is
\[
\begin{aligned}
\mathcal{J}_{\text{A}^2\text{GRPO}}(\theta)
&=
\mathbb{E}_{\{o_i\}_{i=1}^G \sim \pi_{\theta_{\text{old}}(\cdot|q)}}
\Bigg[
\frac{1}{G}\sum_{i=1}^{G}\frac{1}{|o_i|}
\sum_{t=1}^{|o_i|}
\Big(
\min(\varphi_{i,t}A_i,\ \mathrm{clip}(\varphi_{i,t},1-\varepsilon,1+\varepsilon)A_i)
\\
&\quad
-\beta\, m_{i,t}\,(\phi_{i,t}-\log \phi_{i,t}-1)
\Big)
\Bigg],
\end{aligned}
\]
where the KL-like regularization term is multiplied by \(m_{i,t}\), so it applies only to answer tokens [2511.22645].

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 [2511.22645]. 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 \(0\%\) [2511.22645]. GeoZero’s RL stage reverses this behavior.

A key ablation compares training paradigms on UCM and RSVG [2511.22645]. 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 [2511.22645].

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 [2511.22645]. This indicates that the difficulty distribution of RL data is not incidental but constitutive of the method.

Ablations of A$^2$GRPO further show that both the thinking reward and the token mask are required [2511.22645]. Without AMTR, the model tends not to reason. Without masking, reasoning exploration is constrained by the reference model. Full A$^2$GRPO is the only setting that consistently yields both reasoning and strong visual grounding accuracy [2511.22645].

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 \(q_t\) and the total thinking-quality score \(s_t\), while very long or semantically diffuse reasoning can reduce performance [2511.22645]. 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 [2511.22645]. 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 [2511.22645] |
| Visual Grounding | RSVG 50.04%, DIOR-RSVG 79.43%, VRS-VG 73.32% | with RFT [2511.22645] |
| VQA | RSVQA-HR Test1 Presence 91.16%, Compare 91.01% | with RFT [2511.22645] |
| Image Captioning | strong CIDEr results, e.g. UCM-Captions 393.57 | with RFT [2511.22645] |
| XLRSBench / CHOICE | 48.10 / 72.06 average L3 accuracy | with RFT [2511.22645] |

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 [2511.22645]. 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 Acc@0.5 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 Acc@0.5 66.51, Test 68.11, Val gIoU 67.04, Test 69.29** [2605.04451]. 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" [2605.04451]. It addresses the task of mapping a remote-sensing image \(\mathcal{I}\) and a natural-language query \(\mathcal{Q}\) to a reasoning sequence \(c\) and a bounding box
\[
b=[x_1,y_1,x_2,y_2],
\qquad
c,b \sim \pi_\theta(\cdot \mid \mathcal{I}, \mathcal{Q}),
\]
while assuming **no ground-truth boxes during training** [2605.04451].

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 [2605.04451]. RemoteZero therefore replaces external geometric supervision
\[
\mathcal{R}_{ext}(b)=\mathrm{IoU}(b,b_{gt})
\]
with an intrinsic semantic reward
\[
\mathcal{R}_{int}(b)=V(\mathcal{T}(\mathcal{I},b),\mathcal{Q}),
\]
where \(\mathcal{T}\) is a crop operator and \(V\) is an MLLM verifier [2605.04451]. The complete reward adds an area penalty,
\[
r(b,\mathcal{Q})=s-\lambda \max\left(0,\frac{\mathrm{Area}(b)}{\mathrm{Area}(\mathcal{I})}-\tau\right),
\]
to prevent trivial large-box solutions [2605.04451].

Training proceeds via **Group Relative Policy Optimization (GRPO)** using groups of \(K=4\) trajectories per prompt, and crucially the objective contains no \(b_{gt}\) or IoU-to-ground-truth term [2605.04451]. RemoteZero also supports iterative self-evolution: at iteration \(k>0\), the frozen previous policy acts as verifier,
\[
V^{(k)}(\cdot) \leftarrow \pi_{\theta^{(k-1)}}(\cdot),
\]
so the model improves on unlabeled Earth observation data using its own verification signal [2605.04451].

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 [2511.22645].
- **Verification-anchored GeoZero**: supervision only through intrinsic semantic checking of predicted regions, without geometric labels [2605.04451].

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-\(\delta\) intrinsic reward derived from predictive shifts between unassisted and hint-conditioned responses [2605.09959]. 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 [2605.09959, 2511.22645].

GeoZero is also adjacent to GRPO-based optimization work in general. RemoteZero uses GRPO for localization without box supervision [2605.04451], while GeoZero adapts GRPO into A$^2$GRPO by selectively masking KL regularization over answer tokens [2511.22645]. 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 [2511.22645]. Conversely, RemoteZero shows that the same label can extend beyond reasoning traces to spatial endpoint supervision, but still within remote sensing [2605.04451].

## 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 [2511.22645]. The backbone is fixed to Qwen3-VL-8B-Instruct, so backbone sensitivity remains open [2511.22645]. Reinforcement learning is computationally costly, and the definition of "hard" examples depends on a DFM whose own failure modes may bias the RL curriculum [2511.22645]. 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 [2511.22645].

RemoteZero identifies a different set of limitations for GeoZero-style localization. Its reward emphasizes semantic correctness over boundary precision, which explains why Acc@0.5 can exceed supervised baselines while gIoU remains lower [2605.04451]. 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 [2605.04451].

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 [2605.04451]. 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 [2511.22645].

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 [2511.22645, 2605.04451].

Source: https://www.emergentmind.com/topics/geozero