acceptodds
Withdrawn from ICLR 2027

VECPO: View-Equivariant Constrained Policy Optimization for Robust Multi-View Remote Sensing Agricultural Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agricultural multimodal large language models must reason over heterogeneous observations spanning ground-level, unmanned aerial vehicle, and satellite imagery. Although supervised fine-tuning on multi-view data improves domain knowledge, subsequent reinforcement learning faces two core challenges: (i)transformation inconsistency, where predictions become inconsistent under semantics-preserving spatial transformations; and (ii)hidden group regression, where aggregate optimization conceals regressions on underrepresented view–task groups. To address both challenges, we propose View-Equivariant Constrained Policy Optimization (VECPO), a critic-free framework comprising Dual-Granularity Constrained Advantage (DGCA) and the Warm-Start Group Constraint Controller (WGCC). VECPO constructs an identity branch and eligible transformation branches with jointly transformed visual inputs and targets, canonicalizes predictions into a shared reference frame, and assigns correctness-gated equivariance rewards. DGCA couples a source-level relative advantage across all branches with branch-level relative advantages, while WGCC monitors view–task performance against supervised warm-start anchors and supplies group weights that constrain the dual-granularity advantage when a group deteriorates. Experiments on the 13-task AgroMind benchmark, including comparisons with PPO and multiple critic-free policy optimization baselines, show that VECPO achieves strong and competitive performance across diverse agricultural perception and reasoning tasks. Code is available in the supplementary material.

Primary area:
reinforcement learning
Keywords:
Reinforcement learning, Multimodal Large Language Models, Equivariant learning, Multi-view learning, Agriculture, Remote Sensing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZtF7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VECPO, a critic-free policy optimization framework for multi-view agricultural multimodal reasoning. DGCA mixes source-level and branch-level advantages over identity/transformed rollouts with correctness-gated equivariance rewards; WGCC reweights view–task–answer groups against warm-start anchors when they degrade. Experiments on AgroMind compare VECPO with PPO and four critic-free baselines.

Critical strengths

The paper targets two relevant problems: transformation consistency and hidden group regression in multimodal policy optimization.

The empirical setup is reasonable: all RL methods start from the same SFT checkpoint, use a common data partition, and are compared against multiple baselines.

Critical weaknesses

The evidence is too limited to support the main claim. All main RL results use one backbone, one benchmark, and one training seed (42), with no variance or significance tests. Table 1 also contradicts the claim of balanced performance: VECPO is 50.03 vs. PPO 56.14 on SL and 33.38 vs. GRPO 41.92 on VPR, far larger than the stated 1.98 gap.

The ablations do not directly test the proposed mechanisms. DGCA and WGCC are motivated by equivariance and group regression, but Table 2 only reports task accuracy; Figure 4 compares VECPO only with GRPO, not VECPO without DGCA; and WGCC lacks held-out worst-group or group-regression statistics.

Questions for authors

How do you reconcile Table 1 with the claim that the gap on the remaining tasks is at most 1.98, especially on SL and VPR?

Reviewer_xZJ8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Agricultural multimodal tasks span a diverse range of imaging platforms, task objectives, and answer formats. Optimizing solely for an overall response reward guarantees neither that a model will maintain accuracy under spatial transformations nor that performance will be preserved across individual task subsets. To address this, the authors propose incorporating both transformation-based constraints and group-level performance safeguards relative to the initial model.

This paper introduces VECPO, a framework designed for agricultural multimodal reasoning across ground-based, drone, and satellite imagery. The method co-samples original inputs alongside their spatially transformed counterparts. It computes a “correctness-gated” cross-branch reward by mapping targets and applying inverse normalization to the model’s predictions. Specifically, the approach employs a “Dual-Gated Cross-Branch Advantage” (DGCA) that integrates sample-level and branch-level advantages, while a “Weighted Group-Consistency Constraint” (WGCC) adjusts update weights based on performance shifts—relative to an initial anchor checkpoint—within specific “viewpoint–task–answer-type” groups.

Starting from a shared Qwen2.5-VL-7B SFT initialization, VECPO achieves a macro-average accuracy of 49.20% across 13 tasks in the AgroMind benchmark, outperforming both GRPO (46.76%) and SFT (35.85%). The paper further provides component ablation studies, post-transformation accuracy metrics, an analysis of controller training dynamics, and qualitative case studies. All training results are reported based on a single random seed.

Critical strengths
  1. The core motivation is valid and important.
    The paper’s motivation that models should adhere to spatial transformation correspondences is valid and well founded. This issue is practically relevant for agricultural multimodal reasoning, and relying solely on an overall response reward indeed makes it difficult to guarantee transformation consistency.

  2. The task-specific transformation design offers practical value.
    The proposed method distinguishes between answer invariance and spatial equivariance, processes images, relevant prompts, and targets in tandem, and subsequently maps predictions back to a common reference frame. This design is well suited for structured outputs such as coordinates, orientations, and grid locations.

  3. The correctness-gated mechanism is well designed.
    Correctness-based gating prevents the model from being erroneously rewarded for answers that are consistent yet incorrect, thereby coupling cross-branch consistency with answer correctness.

  4. The overall method structure is coherent.
    The method follows a “general–specific–general” (overview–detail–synthesis) pattern, and the overall computational workflow is largely clear and coherent.

  5. The comprehensive training recipe yields broad preliminary performance gains.
    As shown in Table 1, VECPO outperforms the strongest reported RL baseline by 2.44 percentage points and surpasses SFT across all 13 tasks. The use of shared initialization and foundational rewards enhances the interpretability of the comparisons; meanwhile, the superior accuracy observed across three distinct types of spatial transformations further underscores the practical utility of the method.

Critical weaknesses
  1. The motivation that overall improvements can mask localized degradation is valid, but subgroup-level evidence is insufficient.

    • Table 1 does show performance drops on some tasks, but the authors fail to fully demonstrate degradation within the specific subgroups they define.
    • WGCC uses “viewpoint–task–answer-type” groups, yet Table 1 reports only task-level results. The paper does not provide a comprehensive breakdown of degradation across these finer-grained groups, nor does it demonstrate that such degradation is primarily concentrated in low-frequency groups.
    • A minor performance drop observed in a single-seed experiment is insufficient to establish consistent, systemic degradation.
    • “Increasing the weight of difficult groups” is not equivalent to “preserving performance relative to SFT.” The latter requires a reliable SFT baseline, and the current construction of the anchor does not provide such a baseline.
  2. The concept of an “SFT anchor” does not fully align with the algorithm’s description, and the definition of the anchor requires clarification.

    • Section 3.5 characterizes WGCC as a method for preserving performance relative to the SFT warm-start model.
    • However, Algorithm 1 updates the anchor based on observations during the first 32 steps—the same phase in which the policy parameters are also being updated—and the anchor is only frozen after this warm-up period. Appendix D.3 further clarifies that the anchor is derived from an exponential moving average (EMA) of these observation groups.
    • Given this process, the anchor incorporates the performance of the evolving early-stage RL policy and cannot be strictly equated with the performance of a frozen SFT model.
    • The paper also fails to report the sample support for each anchor or to specify how groups that do not appear during the warm-up phase are initialized. While the current evidence supports the plausibility of adaptive reweighting, it is insufficient to confirm that it effectively protects fine-grained subgroups relative to the SFT baseline.
  3. The experiments do not sufficiently disentangle the respective contributions of data augmentation, cross-branch rewards, and the DGCA mechanism.

    • Table 2 does not specify which alternative advantage estimator is used for the “w/o DGCA” ablation, nor which transformation branches, reward components, and controller settings are retained.
    • The paper does not provide a matched GRPO baseline trained with the same transformation branches, nor an experiment isolating the removal of cross-branch rewards. Consequently, the reported results do not clearly attribute performance gains to any specific design component.
    • For discrete tasks with a single canonical correct answer, the condition that both branches are correct inherently implies agreement on the canonical answer; thus, the reward defined in Equation (3) primarily reflects joint correctness across branches.
    • Figure 4 reports only post-transformation accuracy, omitting original accuracy on the same subset and pairwise shifts in predictions—data necessary to distinguish between an overall improvement in capability and a reduction in sensitivity to transformations.
    • Furthermore, reliance on single-seed results limits confidence in the impact of individual components, particularly given the substantial 30.38-percentage-point drop in Accuracy Rate (AR) observed when DGCA is removed.
Questions for authors
  1. DGCA implementation and attribution.

    • How exactly is the advantage calculated for the “w/o DGCA” setting in Table 2? Which transformation branches, reward components, and controller settings are retained?
    • Could you provide a GRPO baseline with transformation-based data augmentation under the same constraints on source samples and response budget?
    • Could you conduct a small-scale targeted ablation to disentangle the contributions of the cross-branch reward, source-level-only rewards, branch-level-only rewards, and mixed-advantage formulations?
  2. Anchor construction and subgroup protection.

    • Are the anchor points derived from a frozen SFT policy or from a policy that is continuously updated during the early stages of training?
    • How many anchor observations are collected for each group, and how are groups that do not appear during the warm-up phase initialized?
    • Where group labels and sample sizes permit, could you report the test-group performance shifts of the full method (and the version without the controller) relative to the SFT baseline—specifically, the maximum degradation and the proportion of degraded groups?
  3. Transformation robustness diagnostics.

    • On the same diagnostic subset used in Figure 4, could you report raw accuracy, the pairwise consistency rate after inverse normalization, and the proportion of cases in which an initially correct prediction becomes incorrect following a transformation?
    • Please further distinguish between spatially sensitive and spatially invariant tasks.
    • Providing uncertainty estimates for these pairwise comparisons—along with validation using a limited number of additional training seeds—would significantly influence how I interpret these results.
Reviewer_CjcMOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Authors propose a VECPO method which a critic-free RL method, containing Dual-Granularity Constrained Advantage (DGCA) and the Warm-Start Group Constraint Controller (WGCC), for agriculture multimodel LLM. The model is initialized using SFT. During RL, author handles transformation inconsistency when the input image is flipped or rotated. Another is hidden group regression, where overall reward goes up when view-task groups get worse. To handle these, authors proposed VECPO which keep identity branch and transformed branch for each transformation example. It introduces equivariance gates which are rewarded when the answers are correct. DGCA gives consistency bonus only when answers from source and transformation are correct. WGCC considers accuracy for each view/task/answer-type against a warm-start anchor and wights the groups that fall behind.

Critical strengths

A) Authors proposed a critic free policy optimization that introduces transformation equivariant learning into multi-view agriculture reasoning.

B) The proposed DGCA combines the source-level comparison across identity and transformed branches with branch-level discrimination to handle transformation inconsistency.

C) Authors analyze, using WGCC, view-task-group performance with respect to warm-start anchors and adapt dual-graduality advantage whenever a group deteriorates providing hidden group-level explicit optimization concerns.

Critical weaknesses

A) The motivation is weak. Author claims to develop a method that can handle flips and rotation. But Agriculture data does not need to consider when to flip or rotate. The agri-VLMs suppose to answer correctly without considering flips/rotation transformations, since the satellite and UAV images are taken from above. Once can say that authors can apply data augmentations to images and the object bounding boxes accordingly. Such data transformations make the paper's motivation weak.

B) From Table 2, removing either DGCA or WGCC, does not beat the second row GRPO on the overall macro average scores.

C) The authors claim that “product uᵢuⱼ opens the gate only when both predictions are correct”. It means that bonus will be given when both are exactly correct. Training on flipped images with flipped boxes and labels will do the same. It does not tell whether VECPO’s design helps at all.

D) It would be better to see results on multiple benchmarks and SOTA comparison with existing agriculture VLM methods.

Questions for authors

Please see identified weakness points and It is better to see the GRPO score with same flipped and rotated data.

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

Related papers

Discussion (0)

Sign in to comment.