GeoZero-Hard in Geospatial RL

• GeoZero-Hard is a hard-sample subset in the GeoZero framework, specifically curated for challenging geospatial multimodal tasks such as scene classification, visual grounding, VQA, and image captioning.
• It is constructed through an automatic, model-dependent hard-sample mining pipeline that balances samples across tasks and isolates cases requiring deep, multi-step spatial inference and fine-grained map-reading.
• By driving policy optimization via A²GRPO, GeoZero-Hard facilitates emergent reasoning, significantly improving performance metrics across remote-sensing benchmarks.

Searching arXiv for the GeoZero paper and closely related reinforcement-learning geospatial reasoning work. GeoZero-Hard is the dedicated hard-sample pool used in GeoZero’s reinforcement learning stage for geospatial multimodal LLMs. In the GeoZero framework, it is paired with GeoZero-Instruct, which comprises 610 K relatively straightforward image–instruction pairs for supervised fine-tuning, whereas GeoZero-Hard contains only the most challenging examples, approximately 20 K samples drawn from four core remote-sensing tasks: scene classification, visual grounding, VQA, and image captioning. Its function is not to provide predefined chain-of-thought supervision, but to stimulate deep reasoning during reinforcement learning by concentrating optimization on cases that require multi-step spatial inference, fine-grained map-reading, and contextual remote-sensing analysis (Wang et al., 27 Nov 2025).

1. Definition and Scope

GeoZero-Hard is defined as the hard-sample subset in GeoZero’s two-stage training design. GeoZero-Instruct is used to teach fundamental geospatial knowledge, including terminology and single-step scene description, while GeoZero-Hard is reserved for cases intended to force the model to engage in deep, multi-step reasoning (Wang et al., 27 Nov 2025).

The targeted challenges are explicitly geospatial. The dataset selects cases requiring multi-step spatial inference, such as counting or comparing objects across distant regions; fine-grained map-reading, such as localizing small targets under confusing backgrounds; and contextual remote-sensing analysis, such as interpreting environmental patterns and scene-level relationships. These examples go beyond read-off tasks and demand an internal chain of thought (Wang et al., 27 Nov 2025).

This division of labor is central to the framework. GeoZero-Hard is not a generic difficulty split; it is a task-balanced hard subset constructed to support reinforcement learning on remote-sensing reasoning problems rather than direct answer imitation. A plausible implication is that the dataset is designed to separate knowledge acquisition from reasoning elicitation, with the latter concentrated in a much smaller but more adversarial training pool.

2. Construction Pipeline

GeoZero-Hard is derived from an automatic hard-sample mining pipeline that begins with GeoZero-Raw, approximately 754 K samples assembled from publicly available SC, VG, VQA, and IC datasets (Wang et al., 27 Nov 2025). The pipeline proceeds in six steps:

1. Assemble GeoZero-Raw from publicly available scene classification, visual grounding, VQA, and image captioning datasets.
2. Train a Data Filtering Model via supervised fine-tuning on GeoZero-Raw.
3. Run one forward pass of the Data Filtering Model on all GeoZero-Raw samples and retain the samples it mispredicts.
4. For each retained candidate, perform 3 stochastic inferences and compute a per-sample difficulty score defined as $1 - (\text{average correctness})$.
5. Rank candidates by difficulty and pick the top portion from each task to ensure 1:1:1:1 task balance.
6. Remove any imagery overlap between GeoZero-Hard and the remaining samples; the remainder becomes GeoZero-Instruct (Wang et al., 27 Nov 2025).

The construction is therefore model-dependent: hardness is determined by failure modes of the Data Filtering Model rather than by manual difficulty annotation. The selected examples all lie in the top decile of error, with difficulty tiers implicitly defined by quantiles of the Data Filtering Model error rate (Wang et al., 27 Nov 2025). This suggests that GeoZero-Hard operationalizes difficulty as persistent model uncertainty under stochastic inference rather than as a purely semantic property of the task.

3. Composition and Annotation Regime

GeoZero-Hard contains approximately 20 000 samples, with an approximately uniform task breakdown: 5 000 scene-classification, 5 000 visual-grounding, 5 000 VQA, and 5 000 image-captioning examples (Wang et al., 27 Nov 2025). The enforced 1:1:1:1 balance is a design constraint of the mining procedure rather than an incidental property of the source corpora.

Component	Description	Scale
GeoZero-Raw	Source pool from public SC, VG, VQA, IC datasets	≈ 754 K
GeoZero-Hard	Hard-sample pool for RL	≈ 20 K
GeoZero-Instruct	Remaining deduplicated data for SFT	≈ 610 K

Ground-truth labels are inherited from the original public benchmarks: class names for scene classification, bounding boxes for visual grounding, reference answers for VQA, and reference captions for image captioning (Wang et al., 27 Nov 2025). No additional manual re-annotation is performed. Quality control instead relies on three mechanisms: use of high-quality public datasets with established benchmarks, multi-round stochastic evaluation in the second-stage filtering to reduce false positives, and image deduplication to avoid leakage between the supervised fine-tuning and reinforcement learning stages (Wang et al., 27 Nov 2025).

Because there is no additional manual re-annotation, GeoZero-Hard preserves the annotation protocols of its constituent benchmarks. A plausible implication is that its novelty lies primarily in sample selection and training role rather than in new human-generated labels.

4. Role in Reinforcement Learning and A$^2$GRPO

In the reinforcement learning stage, GeoZero-Hard is used to compute policy gradients that encourage the model to generate explicit reasoning chains before answering. Training is initialized from the supervised fine-tuned model trained on GeoZero-Instruct, after which Answer-Anchored Group Relative Policy Optimization, denoted A$^2$GRPO, is run on GeoZero-Hard (Wang et al., 27 Nov 2025).

Let $\theta$ denote model parameters, and let $\pi_\theta$ be the policy that outputs a complete response $o = (o_1,\dots,o_n)$ containing both reasoning and answer tokens. The method samples $G$ candidate responses $\{o^i\}_1^G$ under $\pi_{\theta_{\text{old}}}$ (Wang et al., 27 Nov 2025). The reward design has three components:

• The answer reward $r_a \in [0,1]$, a task-specific continuous score measuring answer correctness.
• The answer-modulated thinking reward $^2$0, which ties reasoning quality $^2$1 to answer correctness:

$^2$2

with

$^2$3

• The total reward:

$^2$4

The answer reward is instantiated differently by task: cosine similarity for scene classification and VQA, IoU for visual grounding, and a weighted combination of BLEU, CIDEr, METEOR, ROUGE, and semantic embedding for image captioning (Wang et al., 27 Nov 2025).

The optimization objective is given as:

$^2$5

subject to

$^2$6

Here $^2$7 is the advantage computed from the per-sample rewards $^2$8, and $^2$9 is a thinking mask with $^2$0 for answer tokens and $^2$1 for reasoning tokens (Wang et al., 27 Nov 2025). The mask enforces KL penalty on answer tokens to keep answers near the reference policy, while leaving reasoning tokens unregularized. According to the paper, this prevents collapse to copying a reference chain of thought and fosters diverse, emergent reasoning (Wang et al., 27 Nov 2025).

5. Answer Anchoring and Emergent Reasoning

The answer anchoring mechanism is defined by the interaction between $^2$2, the gating function $^2$3, and the thinking mask $^2$4 (Wang et al., 27 Nov 2025). The answer reward gates the thinking reward: only if $^2$5 exceeds threshold $^2$6 does the model receive credit for $^2$7. Reasoning is therefore rewarded only when it is associated with sufficiently correct answers.

The binary thinking mask separates regularization of answers from exploration of reasoning. Answer tokens remain anchored to the reference policy, whereas reasoning tokens are not directly regularized (Wang et al., 27 Nov 2025). In the paper’s formulation, these design choices together anchor the reasoning process on correctness while encouraging novel internal chains of thought.

This mechanism is presented as a way to obtain reasoning without any predefined chain-of-thought supervision. GeoZero-Hard is the data substrate that makes this possible: because it concentrates hard cases from four task families, policy optimization is repeatedly exposed to samples for which shallow answer generation is insufficient. The reported result is universal emergent reasoning capabilities across diverse geospatial tasks (Wang et al., 27 Nov 2025). This suggests that dataset hardness and reward structure are intended to be complementary rather than separable components.

6. Evaluation, Effects, and Limitations

Evaluation covers classical tasks on unseen test sets, challenging benchmarks, and emergent reasoning diagnostics (Wang et al., 27 Nov 2025). The classical tasks include scene classification accuracy on UCM, AID, RESISC-45, EuroSAT, SIRI-WHU, and WHU-RS19; visual grounding [email protected] on RSVG, DIOR-RSVG, and VRS-VG; VQA Presence/Compare accuracy on RSVQA-HR Test1/Test2; and captioning metrics BLEU-4, CIDEr, METEOR, and ROUGE-1/L on UCM-Caps, Sydney-Caps, NWPU-Caps, and RSICD (Wang et al., 27 Nov 2025). Challenging benchmarks include XLRS-Bench, evaluated by average L3 subtask accuracy, and CHOICE, evaluated by average accuracy. Emergent reasoning diagnostics include Thinking Activation Rate and the correlation of reasoning length and thinking-quality score $^2$8 with answer accuracy (Wang et al., 27 Nov 2025).

The quantitative improvements reported for GeoZero with reinforcement fine-tuning on GeoZero-Hard are substantial. Scene classification average accuracy increases from approximately 93.8% to 95.5%; visual grounding average [email protected] from approximately 43% to 67%; VQA presence/compare from approximately 75%/84% to 91%/91%; and image captioning CIDEr on UCM-Caps from approximately 280 to 395, with similar gains on other datasets (Wang et al., 27 Nov 2025). On challenging benchmarks, XLRS-Bench improves from 47.2% to 48.1%, and CHOICE from 71.7% to 72.1% (Wang et al., 27 Nov 2025).

Ablations identify GeoZero-Hard as necessary for the emergence of reasoning in this framework. Replacing GeoZero-Hard with random GeoZero-Raw samples abolishes reasoning, producing 0% activation, and greatly degrades visual grounding performance, with [email protected] dropping by approximately 30 points (Wang et al., 27 Nov 2025). Additional ablations show that removing answer-modulated thinking reward or the thinking mask eliminates emergent reasoning or reduces accuracy (Wang et al., 27 Nov 2025).

Qualitative examples illustrate the intended behavior. On a DIOR-RSVG instance, “Locate the cluster of small white structures,” the model first describes “a tightly grouped set of bright rectangular shapes near the top-left, with consistent spacing…these resemble building rooftops…” and then outputs precise bounding-box coordinates. On a VQA example, “How many boats are docked?”, it reasons “I count three elongated objects adjacent to the shore…each with a distinct wake pattern.” and answers “3.” The paper states that such coherent, stepwise reasoning trajectories appear consistently across scene classification, visual grounding, VQA, and image captioning tasks, providing human-interpretable justifications (Wang et al., 27 Nov 2025).

The limitations are also explicit. Reasoning collapse is reported on the very hardest cases: when IoU $^2$9, reasoning chains become repetitive or incoherent (Wang et al., 27 Nov 2025). The current scope is restricted to four tasks, and extending the design to multi­temporal, hyperspectral, or vector-map reasoning remains open. Efficiency is another concern, as reasoning length sometimes exceeds necessity, described as overthinking. The paper also notes possible bias and coverage issues in automatic hardness selection, including overrepresented geographic regions (Wang et al., 27 Nov 2025).

Taken together, GeoZero-Hard is a principled, largely automatic hard-sample pool for reinforcement learning in geospatial multimodal reasoning. Its defining characteristics are task-balanced hardness mining, inheritance of benchmark annotations without manual chain-of-thought curation, and use within A$\theta$0GRPO to tie reasoning credit to answer correctness while leaving reasoning trajectories free to diversify (Wang et al., 27 Nov 2025).