acceptodds
Under review as a conference paper at ICLR 2027

GeoVR: Structured Execution Rewards for Reinforcement Learning in Geospatial Code Generation

Abstract

Large language models are increasingly used to generate geospatial analysis code, yet executable programs can still be analytically wrong: an incorrect data source, temporal range, coordinate reference system, or workflow often fails silently. Existing reinforcement learning with verifiable rewards (RLVR) for code relies on unit tests or execution outcomes, which cannot detect such errors. We introduce GeoVR, an RLVR framework for geospatial code generation built on Google Earth Engine (GEE). Its core reward, Hierarchical Schema-grounded Verifiable Reward (HSVR), extracts type-aware signatures from GEE execution artifacts (object type, bands, CRS, spatiotemporal extent, and verifiable statistics) and checks required analytical operations in the program AST, yielding four deterministic components: Execution, Type, Schema, and Method. A hierarchical gate activates Schema and Method rewards only when the primary result passes task-level verification with the correct output type, preventing incorrect programs from accumulating reward through partially satisfied constraints. GeoVR is optimized with GRPO and leverages structured execution errors to repair failed rollouts. On AutoGEEval++, GeoVR achieves pass@1 of 90.22%, 88.76%, and 67.82% on Atomic, Combined, and Theme tasks, outperforming CodeGEEnius by up to 6.38 points and a controlled Binary-RLVR baseline by 3.73–10.35 points. These results demonstrate that execution artifacts themselves can serve as effective verification feedback for reinforcement learning, enabling models to move beyond generating merely "executable programs" toward generating "task-correct programs that satisfy domain constraints." Data and code are available at: https://figshare.com/s/03091d54357e8a152e7a

Primary area:
reinforcement learning
Keywords:
Large language models, reinforcement learning with verifiable rewards, code generation, structured execution feedback, reward design, domain-specific AI, Google Earth Engine, geospatial artificial intelligence

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UrucOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GeoVR studies reinforcement learning for Google Earth Engine code generation. It constructs a hierarchical reward from execution correctness, output type, output schema, and required analytical operations. Schema checks use execution artifacts such as bands, coordinate reference systems, and spatiotemporal properties, while method checks use the program AST. The method combines these rewards with GRPO and execution-feedback repair. Evaluation on AutoGEEval++ includes comparisons against binary-reward RLVR and ablations of the reward components.

Critical strengths
  1. The paper identifies a meaningful domain problem: successful execution alone does not establish that a geospatial program satisfies the requested analytical constraints. Turning observable execution artifacts into supervision is practically useful.

  2. The component ablations show that schema checks, method checks, and hierarchical gating each contribute to benchmark performance. This supports the value of the domain-specific verifier rather than relying only on comparisons with general-purpose models.

Critical weaknesses
  1. HSVR looks largely like a domain-specific test suite turned into an RL reward. Checking types, metadata, and analytical constraints is useful, but the paper does not clearly explain what new learning principle this introduces.

  2. The hierarchy is a fairly straightforward way to enforce prerequisites, and the main comparison changes both the reward and the repair mechanism. This makes it difficult to tell how much of the contribution comes from a new reward design rather than careful engineering for GEE.

Questions for authors
  1. What is the precise distinction between HSVR and a conventional domain-specific test suite whose assertions check output type, schema, metadata, and analytical operations? Which part of the contribution changes the RL methodology, rather than expanding the set of properties verified by execution-based tests?

  2. Beyond a conventional gated or constrained reward, what new principle does the Execution → Type → {Schema, Method} hierarchy introduce? Please explain what the gating ablation establishes about this principle, as opposed to the usefulness of enforcing prerequisites in this particular verifier.

  3. Which components of GeoVR constitute a reusable method beyond GEE, and which require task- or platform-specific specification? Please clarify how required operations and schema constraints are derived without simply enforcing the reference implementation, and what existing evidence supports accepting functionally equivalent alternative workflows.

3 comments
AuthorsOpenReview ↗

Clarified HSVR vs. test suites and the gating principle, added reward/repair ablation

We thank the reviewer.

Weakness 1 and Question 1. HSVR differs from a domain test suite in three ways. (1) What is verified. A test suite asserts output values for fixed inputs. A GEE output is a geographic object, and the same area value can come from a wrong period, a wrong projection, or a table missing the region field, each making the result unusable. HSVR extracts signatures by object type: bands, CRS, resolution, and spatiotemporal extent for images, and fields, geometry type, and extent for vectors. (2) How checks are combined. Test assertions are independent and at most summed. HSVR applies Schema and Method only after the primary result and output type pass, so a wrong-type program cannot earn reward from coincidentally correct bands or CRS. (3) How constraints are set. Tests derived from a reference assert all its properties and force reproduction of one implementation. HSVR activates only properties the instruction requires and can verify, and marks the rest unconstrained (p. 5, L217–220). If no output projection is requested, UTM and WGS84 receive the same reward, as do normalizedDifference and a band expression for NDWI.

The new learning principle is to build rewards from geospatial constraints in execution artifacts, ordered by their dependencies. This structure sets the ranking within each GRPO group. Binary reward saturates once most candidates pass, and within-group advantages collapse. In Fig. E2, Binary-RLVR's maximum advantage keeps decreasing while GeoVR keeps a broader positive range (p. 24, L1286–1289). With close Execution Rewards, 0.910 vs. 0.945, Theme pass@1 differs by 10.35 points. Figs. G1–G3 show executable programs with correct primary results that HSVR separates and binary reward cannot. Figures: G1 · G2 · G3 · E2

Weakness 2. Table 3 crosses reward type with repair (p. 6, L308–310).

Table 3 (%)Binary K=0Binary K=2HSVR K=0HSVR K=2
Atomic@186.4988.5589.2390.22
Combined@178.8684.0086.4888.76
Theme@157.4758.6263.2267.82
GeoCodeEval@180.6983.1783.6686.63

The two K=0 runs share the GEE environment, tasks, GRPO configuration, and execution count. Only the reward differs, and HSVR leads by 2.74, 7.62, and 5.75 pass@1 points and 2.97 GeoCodeEval points. HSVR at K=0 also beats Binary at K=2. Repair adds 0.99, 2.28, and 4.60 points, so both contribute, the reward more. The checks are defined per object type rather than per task, and the gain carries over to Llama-3.2-3B, where GeoVR exceeds Binary-RLVR by 4.59 points on Theme and 3.96 on GeoCodeEval. The hierarchy's order follows the geospatial workflow, and removing it with identical checks costs 8.05 Theme points (Question 2).

Question 2. Reward organization should follow the dependencies among task constraints. In a geospatial workflow, the output type decides which operations, such as spatial joins, zonal statistics, and export, are possible, and bands, CRS, extent, and fields are attributes of an object of the correct type. HSVR w/o Gating keeps identical checks and training conditions (p. 8, L418–420) and removes only these dependencies, which amounts to a test-suite-style sum. It loses 1.51, 3.81, and 8.05 pass@1 points on Atomic, Combined, and Theme (Table 4), so the drop grows with the number of constraints. Output Type Mismatch rises from 2.33% to 3.26% and Parameter Error from 1.83% to 2.37% (Table 5): without gating, wrong-type programs earn partial reward, and the policy errs more on prerequisites. The same dependencies exist in ArcPy and PyQGIS.

Question 3. Reusable components are the task-specification format, gated scoring, and GRPO training. The format covers output type, dataset, extent, time range, structural properties, operations, prohibited patterns, and unconstrained marks. Platform-specific components are signature extraction, operation rules, and error classification; for example, ArcPy reads the spatial reference and cell size from a raster object. Each task's specification is built from the instruction and the reference program's signature and AST. The reference supplies candidate properties, and only those the instruction requires are constrained. For a July 2022 NDWI median composite, region, period, NDWI, and median compositing are constrained, while the reference's cloud threshold, projection, and export scale are not. Automatic consistency checks pass 96.8% of specifications; failures are corrected or removed. Manual inspection of 200 random tasks found 189 (94.5%) consistent and not excluding reasonable alternatives (p. 27, L1439–1442).

AuthorsOpenReview ↗

Follow Up on Our November 16 Response

Dear Reviewer Uruc,

As the discussion period is approaching its deadline, we would like to kindly follow up on our response posted on November 16.

In response to your concerns, we clarified how HSVR differs from conventional test suites and explained its methodological contribution. We also elaborated on the rationale behind hierarchical gating and provided the corresponding ablation results. Additionally, we clarified which components of GeoVR are reusable across geospatial platforms and how our task specifications accommodate functionally equivalent implementations. We further added controlled reward and repair ablations to distinguish their respective contributions.

These clarifications and results are included in our revised manuscript. We would appreciate it if you could revisit our response. If you have any remaining questions or concerns, please let us know before the discussion closes, if possible. We would be happy to provide further clarification.

Thank you very much for your time and consideration.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer Uruc,

We apologize for following up once more as the discussion period is drawing to a close.

We understand that you may be very busy, and we truly appreciate the time and effort you have devoted to reviewing our paper.

We have made our best effort to address your concerns through detailed explanations and additional controlled experiments in our November 16 response. We would be very grateful if you could kindly take another look at our response when you have a moment.

If you find that our clarifications and additional evidence help resolve your concerns, we would greatly appreciate it if you might consider revisiting your overall assessment. We would also be happy to address any remaining questions.

Thank you again for your patience, thoughtful feedback, and valuable time.

Reviewer_yV1GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents GeoVR, a reinforcement learning (RL) framework for improving foundation models on geospatial code generation for Google Earth Engine (GEE). The authors argue that rewards based only on static correctness or successful execution of the generated JavaScript are too coarse for this domain. They propose a Hierarchical Schema-grounded Verifiable Reward (HSVR). Beyond execution success, HSVR inspects the artifacts produced by running the program and checks their type, schema, and method-level correctness. These signals are combined through hierarchical gating. Applied to Qwen2.5-7B-Instruct, GeoVR substantially improves pass@1 on AutoGEEval++ (up to +49.43 points over the base model). The resulting 7B model also outperforms strong closed-source models such as GPT-5.5 and Claude Sonnet 4.5.

Critical strengths
  • Sound, well-grounded approach with empirical gains. RL with verifiable, execution-derived rewards is an established framework for improving code generation. Adapting it to the GEE domain is a natural choice. The gains are large: up to 49.43 pass@1 points over the off-the-shelf backbone.

  • Well-motivated reward design. HSVR’s hierarchical gating (execution → type → schema → method) follows the natural dependency structure of geospatial program correctness. The ablations show, both intuitively and empirically, that this gated combination beats alternative ways of aggregating the components.

  • Strong results against closed-source models. A 7B open model trained with GeoVR outperforms proprietary models, including GPT-5.5 and Claude Sonnet 4.5. This shows that domain-specific training with verifiable rewards can close, and even reverse, the gap with much larger general-purpose models.

Critical weaknesses

W1. Closed-source baselines may be under-optimized. The paper does not say how much prompt engineering the closed-source baselines received. These models know JavaScript well, but GEE programming depends on domain knowledge: the API surface, the ee.* object types etc. Much of this can be supplied in context. The authors should report results with domain-adapted prompts, for example by adding API documentation or a retrieved operator knowledge base such as the one built by GEE-Ops.

W2. Missing training-free baselines that use the same data. GeoVR’s improvement comes with non-trivial training cost. A fair head-to-head comparison would let closed-source models use the same training corpus without training. For example, retrieve the top-K training instructions most similar to each test query and give them, with their reference programs, as in-context exemplars. Without this, it is hard to tell whether the gains come from RL itself or from access to in-domain data.

These experiments (W1 and W2) are needed to show that fine-tuning the underlying model is necessary for competitive GEE code generation, rather than other viable routes.

W3. Execution-feedback repair is a confound. (a) Do the closed-source baselines get the same execution-feedback-based repair loop as GeoVR at inference time? If not, the comparison is not like-for-like. (b) An ablation is missing in which GeoVR is trained without execution-feedback repair. This would separate how much of the gain comes from the model learning to repair its programs, and how much from RL teaching it to generate correct programs on the first attempt.

W4. Generality across backbones. All experiments use Qwen2.5-7B-Instruct. It is unclear whether GeoVR transfers to other model families (e.g., Llama, Mistral, Gemma), other scales (e.g., 1.5B/3B and 14B/32B), or code-specialized backbones (e.g., Qwen2.5-Coder). Results on at least one more family and one more scale would make the generality claims much stronger.

W5. Missing quantitative comparison with domain-specific prior work. The paper should compare on AutoGEEval++ against existing geospatial code generation systems, notably GeoCode-GPT, GeoColab, and GeoCogent.

W6. Benchmark scope. All results are on a single benchmark. Evaluation on at least one other GEE or geospatial code benchmark (e.g., the GeoCode-Bench/GeoCode-Eval from "Hou et al. 2024, Can Large Language Models Generate Geospatial Code?") would show that the gains are not specific to one benchmark.

Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Added evaluation prompts, no-repair ablation, Llama-3.2-3B results, domain-specific baselines, and GeoCodeEval

We thank the reviewer.

W1. All models, including GeoVR, receive the same benchmark prompt, now in Appendix F.3 (Fig. F3, Fig. F4). GPT-5.5 and Claude Sonnet 4.5 match GeoVR on Atomic, which tests individual APIs and ee.* types, so API knowledge is not their bottleneck. They trail by 13.80 and 11.50 points on Theme, where correctness depends on event dates, cloud masking before compositing, CRS, and statistics fields (Table 1). Operator documentation does not encode these constraints; HSVR supervises them. GeoCogent adds planning and operator retrieval, the kind of knowledge base suggested, to the same backbone, yet stays far below GeoVR. Training supplies this knowledge: from GeoVR-SFT (44.83% Theme), Binary-RLVR reaches 57.47% and HSVR 67.82%.

W2. Binary-RLVR and GeoVR share the SFT checkpoint, RL tasks, GRPO configuration, and rollout budget (p. 6, L303–306), i.e., identical data, yet GeoVR leads by 3.73, 9.90, and 10.35 points (Table 2). On the same backbone, GeoCogent and GeoColab raise Theme by at most 8.05 points over the base model, while SFT plus HSVR reaches 67.82%. The two routes are complementary and can be combined.

W3. (a) No model uses repair at inference (p. 7, L352–353); repair is used only in GeoVR's training. (b) We added training without repair (p. 6, L308–310):

Table 3 (%)Binary K=0Binary K=2HSVR K=0HSVR K=2
Atomic@186.4988.5589.2390.22
Combined@178.8684.0086.4888.76
Theme@157.4758.6263.2267.82
GeoCodeEval@180.6983.1783.6686.63

At K=0, HSVR beats Binary by 2.74, 7.62, 5.75, and 2.97 points and also beats Binary at K=2, so the reward alone improves first attempts. Repair adds 0.99, 2.28, and 4.60 points, since GEE errors are diagnostic and repaired rollouts become candidates HSVR can score (p. 8, L408–414).

W4. We repeated all training on Llama-3.2-3B (p. 7, L368–372). GeoVR leads Binary-RLVR by 0.25, 0.76, 4.59, and 3.96 points and the SFT model by 1.06, 9.81, 10.34, and 6.43 points, and as on Qwen2.5-7B the gain grows with task complexity.

Table D6 @1 (%)BaseSFTBinary-RLVRGeoVR
Atomic76.7182.8683.6783.92
Combined32.8665.7174.7675.52
Theme4.6042.5348.2852.87
GeoCodeEval44.5567.3369.8073.76

W5. Table 1 adds GeoCode-GPT, GeoCogent, and GeoColab; the latter two run their own pipelines on Qwen2.5-7B-Instruct (p. 26, L1393–1397). GeoVR leads them by 41.38 to 51.73 Theme points. These systems add knowledge at generation time and HSVR at training time, so the two are complementary.

Table 1 @1 (%)AtomicCombinedThemeGeoCodeEval
Qwen2.5-7B-Instruct80.3745.2418.3972.28
GPT-5.590.0679.1454.0280.20
Claude Sonnet 4.588.7586.3856.3281.68
GeoCode-GPT87.6466.9521.8472.77
GeoCogent72.4251.7116.0977.72
GeoColab87.7668.9526.4478.22
CodeGEEnius89.9582.3862.0779.70
GeoVR90.2288.7667.8286.63

W6. We evaluated all 202 GEE JavaScript tasks of GeoCodeEval from Hou et al. (p. 7, L349–351), scored with the GeoCode-GPT criterion (Hou et al., 2025e). GeoVR leads at every k, 4.95 points above Claude Sonnet 4.5 and 5.94 above Binary-RLVR at @1. GeoCodeEval does not rely on execution, so the gain is not tied to our verifier.

Table D5 (%)@1@3@5
Qwen2.5-7B-Instruct72.2881.6883.66
GPT-5.580.2088.6191.58
Claude Sonnet 4.581.6889.6091.58
GeoCode-GPT72.7784.6587.13
GeoCogent77.7283.1785.64
GeoColab78.2287.6289.60
CodeGEEnius79.7086.6389.11
GeoVR-SFT78.7186.6388.61
Binary-RLVR80.6988.1289.60
HSVR w/o Schema82.6789.1191.58
HSVR w/o Method81.1989.6091.58
HSVR w/o Gating82.1889.6091.58
GeoVR86.6393.0794.55

I appreciate the authors taking the effort to conduct additional experiments on (a) the Llama-3.2-3B backbone, (b) GeoCode-GPT, GeoCogent, and GeoColab baselines, and (c) the GeoCodeEval benchmark. While these significantly strengthen the paper, my concern about insufficient optimisation of closed-source baselines such as GPT-5.5 and Claude Sonnet 4.5 remains.

I am not yet convinced, given access to in-domain data, that optimising closed-source models using methods like Training-free GRPO [1] is less viable than spending the compute on high-latency training associated with GeoVR. This will surely change their input prompt compared to Figures F3 and F4, but does optimising the prompt context yield better performance?

Answers to questions such as whether optimised closed-source baselines still lag behind GeoVR in performance, or whether optimising them to achieve high performance is costly, are missing. Thus, I would like to maintain my current score.

[1] Training-Free Group Relative Policy Optimisation, 2025

AuthorsOpenReview ↗

Added In-Domain Knowledge Retrieval Experiments for Closed-Source Models

We thank the reviewer for recognizing our additional experiments and for suggesting the use of in-domain data to improve closed-source baselines. We further evaluated GPT-5.5 and Claude Sonnet 4.5 with retrieved examples from the full GeoVR training corpus. The results show that retrieval improves both models, but they still trail GeoVR on Combined, Theme, and GeoCodeEval.

Retrieval setup. The knowledge base holds all 176,897 training samples of GeoVR, namely 173,057 SFT instruction–program pairs and 3,840 RL training tasks. Each sample keeps its full instruction and reference GEE JavaScript program without chunking. Instructions are embedded with bge-base-en-v1.5 into 768-dimensional vectors, and retrieval is exact cosine search over the whole corpus. For each test task, the 3 or 5 most similar instruction–program pairs are placed before the task in order of similarity, and the prompt asks the model to adapt them to the data source, region, period, and output of the current task. All models receive the same retrieved examples, and the top-3 examples are the first three of the top-5 examples. Training-free GRPO distills experience from the model's own rollouts into the context, while this setup gives the model the reference programs of the most similar in-domain tasks directly.

Results.

ModelRetrievalAtomic@1 (%)Combined@1 (%)Theme@1 (%)GeoCodeEval@1 (%)
GPT-5.5None90.0679.1454.0280.20
Top-390.2481.4357.4781.19
Top-590.5982.2960.9282.18
Claude Sonnet 4.5None88.7586.3856.3281.68
Top-389.8087.1458.6284.16
Top-590.0687.6259.7783.66
GeoVRNone90.2288.7667.8286.63
Top-391.3889.3368.9788.61
Top-591.7389.8168.9789.11

Bold values mark the best result in each column.

Top-5 retrieval raises Theme by 6.90 points for GPT-5.5 and 3.45 points for Claude Sonnet 4.5. Even with examples drawn from the full GeoVR training corpus, both closed-source models remain below GeoVR without retrieval on Combined, Theme, and GeoCodeEval, with gaps of 1.14, 6.90, and 2.47 points, respectively. On Theme, GPT-5.5 and Claude Sonnet 4.5 reach 60.92 and 59.77 with top-5 retrieval, while GeoVR reaches 67.82 without retrieval. GeoVR also benefits from retrieval. Top-5 retrieval raises it by 1.51, 1.05, 1.15, and 2.48 points to 91.73, 89.81, 68.97, and 89.11, and under the same retrieval GeoVR leads the best closed-source model by 1.14, 2.19, 8.05, and 5.45 points. These results show that retrieval augmentation and HSVR training are complementary.

Cost. One full evaluation of GPT-5.5 costs 37.3, 51.0, and 58.3 USD with no, top-3, and top-5 retrieval, and one of Claude Sonnet 4.5 costs 23.5, 33.5, and 38.9 USD. Top-5 retrieval adds 21.0 and 15.4 USD to each evaluation. GeoVR is trained once in 43.69 hours on 4 A800 GPUs, as reported in Appendix H.3, and runs locally without third-party API calls.

Appendix D.6 now describes the retrieval setup and reports these results in Table D7, and Section 6.1 summarizes them.

We hope these additional results address your remaining concerns and would appreciate it if you could reconsider your assessment in light of the new evidence. We welcome any further questions.

Reviewer_Jzg5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoVR, an RLVR framework for geospatial code generation using Google Earth Engine. It introduces a hierarchical reward based on execution outcomes, output structure, and analytical constraints, combined with execution-feedback-based code repair. Experiments on AutoGEEval++ show improvements over binary-reward RLVR and existing geospatial code generation baselines.

Critical strengths

This paper addresses a practical limitation of binary execution rewards in geospatial code generation. The proposed reward incorporates domain-specific constraints that execution outcomes alone may overlook. The results demonstrate the practical value of structured verification for geospatial code generation, particularly on tasks involving multiple analytical operations.

Critical weaknesses

The proposed framework combines existing RLVR techniques with domain-specific verification rules. While the application is practical, the paper offers limited methodological advances or generalizable insights into RLVR.

The main comparison changes both the reward design and execution-feedback repair, with different execution budgets. Although component ablations are provided, the experiments do not adequately isolate the contributions of these factors. More controlled comparisons, including reward-only, repair-only, and compute-matched baselines, are needed to substantiate the claimed improvements.

I consider this a practical application of existing RLVR techniques to geospatial code generation, with limited methodological novelty. My main concern is the experimental rigor, as the current comparisons do not sufficiently isolate the contributions of the proposed components. I lean toward weak rejection, but would reconsider my assessment if more controlled ablations established the benefits of the reward design independently of execution-feedback repair and additional computation.

Questions for authors

Could the authors provide controlled ablations that isolate the effects of HSVR and execution-feedback repair? In particular, comparisons with reward-only, repair-only, and compute-matched binary RLVR baselines would help establish whether the improvements come from the proposed reward design.

4 comments
AuthorsOpenReview ↗

Added controlled reward/repair ablations and clarified the methodological contribution

We thank the reviewer. All new results are in the revised paper.

W1. Methodological contribution. The problem comes from geospatial analysis itself. GEE hosts global remote-sensing data, and flood assessment, water extraction, vegetation monitoring, and land-cover and urban change analysis all run on it. Generating GEE programs from natural language lets geoscientists and emergency managers who do not program run such analyses directly. General coding ability is not enough: a GEE program is correct only if its data source, time range, CRS, bands, spatial extent, and workflow are all right. When these constraints are violated, GEE usually raises no error, and the result enters area statistics, change detection, or disaster assessment in a normal-looking form, which is often costlier than a crash.

Appendix Figs. G1–G3 show three programs whose primary result matches the reference. The first computes the correct flooded area but returns an ee.Dictionary, unusable for administrative aggregation, spatial joins, or export. The second returns an ee.FeatureCollection without region_id and unit, and its event_date is the pre-event date. The third replaces the required median composite with first(); one image remains after filtering, so its NDWI equals the reference. Binary reward gives all three full credit. GeoVR's contribution is to turn these constraints into verifiable rewards.

HSVR extracts execution signatures by output type: band set and order, data types, CRS, resolution, extent, and temporal metadata for images, and property schema, geometry type, extent, and feature count for vectors. Cloud masking, reducer choice, and operation order cannot be read from the output, so the Method reward checks them on the AST. The signature captures what object is produced, and the AST captures how.

These constraints change GRPO dynamics. Later in training most candidates execute, binary reward saturates, and within-group advantages shrink. In Fig. E2, Binary-RLVR's maximum advantage keeps decreasing, whereas GeoVR keeps a broader positive range (p. 24, L1286–1289). Validation Execution Rewards are close, 0.910 vs. 0.945, yet pass@1 differs by up to 10.35 points (p. 8, L403–405); the margin comes from CRS, spatiotemporal, field, and method distinctions. This arises whenever correctness goes beyond execution, so it is not specific to GEE.

Gating mirrors the order of constraints in a geospatial workflow: if the primary result or output type is wrong, checking bands, CRS, and fields is meaningless (p. 8, L431–432). Removing Schema, Method, or gating lowers Theme pass@1 from 67.82% to 58.62%, 59.77%, and 59.77% (Table 4), and removing gating raises Output Type Mismatch from 2.33% to 3.26% (Table 5), so the dependency structure itself affects learning.

The design carries over to ArcPy, PyQGIS, and PostGIS SQL, whose programs also involve extents, CRS, fields, and time and also fail silently. The specification format, the approach of extracting signatures by object type, gated scoring, and GRPO procedure are reused; only the execution interface, property extraction, and error rules are reimplemented.

Figures: G1 · G2 · G3 · E2

W2 and Q1. Controlled ablations. We added four runs crossing Binary and HSVR with no repair (K=0) or up to two repair rounds (K=2) (p. 6, L308–310). HSVR at K=0 is reward-only, Binary at K=2 is repair-only, and the two K=0 runs are compute-matched: both use only the 61,440 initial rollouts as GEE executions, and HSVR's local checks take about 76 s in total (p. 32, L1720–1723).

Table 3 (%)Binary K=0Binary K=2HSVR K=0HSVR K=2
Atomic@186.4988.5589.2390.22
Combined@178.8684.0086.4888.76
Theme@157.4758.6263.2267.82
GeoCodeEval@180.6983.1783.6686.63

At equal compute, HSVR exceeds Binary by 2.74, 7.62, and 5.75 pass@1 points and 2.97 GeoCodeEval points, and it also beats Binary at K=2 on all four metrics with fewer executions, so the gain comes from the reward itself. With HSVR fixed, training-time repair adds 0.99, 2.28, and 4.60 points, so repair also contributes, as GEE errors such as inaccessible assets, projection errors, or type mismatches are diagnostic.

Overall. Without repair and at equal GEE executions, HSVR outperforms Binary, and it outperforms Binary with repair as well, so the reward's advantage does not depend on repair or extra computation. Methodologically, HSVR turns geospatial constraints into verifiable rewards that keep providing within-group signal after binary reward saturates. We hope this addresses your concerns.

AuthorsOpenReview ↗

Follow Up on Our November 16 Response

Dear Reviewer Jzg5,

As the discussion period is approaching its deadline, we would like to kindly follow up on our response posted on November 16.

Regarding your concerns, we clarified the methodological contribution of HSVR, including how it incorporates geospatial constraints beyond binary execution success and the role of hierarchical gating. We also added four controlled ablation settings comparing binary reward and HSVR with no repair or up to two rounds of execution-feedback repair. The results show that HSVR outperforms binary reward even without repair, helping distinguish the contribution of reward design from execution-feedback repair.

These results and clarifications are included in our revised manuscript. We would appreciate it if you could revisit our response. If you have any remaining questions or concerns, please let us know before the discussion closes, if possible. We would be happy to provide further clarification.

Thank you very much for your time and consideration.

Thank you for the responses and additional experiments. The controlled ablations address some of my concerns regarding experimental rigor. I still find the methodological contribution and the insights offered to the broader RLVR literature limited. I will therefore maintain my original overall rating.

AuthorsOpenReview ↗

Further Clarification and Appreciation

Dear Reviewer Jzg5,

Thank you very much for taking the time to review our additional experiments and for acknowledging that the controlled ablations have addressed some of your concerns regarding experimental rigor. We sincerely appreciate your careful consideration and fully understand that you still have reservations about the methodological contribution.

We are particularly grateful for the clear guidance you provided in your initial review, where you wrote

My main concern is the experimental rigor, as the current comparisons do not sufficiently isolate the contributions of the proposed components. I lean toward weak rejection, but would reconsider my assessment if more controlled ablations established the benefits of the reward design independently of execution-feedback repair and additional computation.

We took your suggestion very seriously and made our best effort to address this concern through additional controlled experiments.

As reported in Table 3, we added four settings that separately examine reward design and execution-feedback repair. With repair disabled, HSVR outperforms binary reward by 2.74, 7.62, 5.75, and 2.97 points on Atomic, Combined, Theme, and GeoCodeEval, respectively. Both settings use the same number of initial rollouts and GEE executions. The comparisons with repair enabled further help distinguish the contributions of reward design and repair.

We hope these results provide more convincing evidence that the improvements from HSVR are not dependent on execution-feedback repair or additional GEE executions.

We would also like to offer one further clarification regarding the broader methodological contribution to RLVR.

Beyond the domain-specific verification rules, our main insight concerns the reward signals used during RLVR training. As more generated programs execute successfully, binary execution rewards tend to saturate, reducing the distinctions among candidates within each GRPO group. As shown in Fig. E2, Binary-RLVR exhibits diminishing advantage signals during training, while HSVR maintains a broader positive advantage range.

By extracting structured information from execution artifacts and organizing rewards according to constraint dependencies, HSVR can continue to distinguish programs that execute successfully but fail to satisfy the intended requirements. We hope this observation offers a useful perspective on reward design for RLVR, particularly for tasks involving structured outputs where execution success alone is insufficient.

We understand that these explanations may not fully resolve your concerns about methodological novelty, and we sincerely respect your perspective. Nevertheless, given that we have made substantial efforts to address the experimental concerns highlighted in your initial review, we would be truly grateful if you might be willing to reconsider your overall assessment in light of the new evidence.

We realize that we are asking for your consideration once again, and we greatly appreciate your patience and the time you have already devoted to reviewing our work.

Thank you again for your thoughtful feedback and for your efforts in helping us improve the quality of our paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.