acceptodds
Withdrawn from ICLR 2027

GeoReflex: Self-Correcting Active Perception for Ultra-High-Resolution Remote Sensing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Ultra-high-resolution (UHR) remote-sensing understanding requires integrating global context and local detail within a limited visual budget. Active-perception methods acquire evidence through dynamic cropping, but incorrect crops can leave subsequent exploration and reasoning dependent on incomplete observations. We introduce GeoReflex, an active-perception framework that uses reasoning to examine and correct multi-step visual observations. It connects evidence diagnosis with observation adjustment: when evidence is insufficient, the model can restore the parent view and use its diagnosis to redirect exploration. To learn this behavior, we construct GeoReflex-48K with examples and combine the capabilities of a Generator, Reflector, and Refiner in a single model through supervised fine-tuning and reflection-efficient reinforcement learning, teaching when to reflect and how to correct. Across three remote-sensing benchmarks, GeoReflex exceeds the strongest compared external baselines on RSHR-Bench and XLRS-Bench by and points, respectively, demonstrating the value of recovery-oriented post-training for UHR understanding. Our anonymous repository is available at https://anonymous.4open.science/r/GeoReflex-CB14/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language Models, Active Perception, Self-Correction, Reinforcement Learning, 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_5KiHOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Authors present a GeoRlex to handle UHR remote-sensing understanding which reasons to examine and corrects using multi-step visual observation. To do so, they curate a dataset called GeoRelex-48K to include the generator, reflector and refiner capabilities through SFT and RL, guiding the model when to reflect and how to correct. Authors evaluated their method over 3 benchmarks showing better scores.

Critical strengths

A) Authors proposed a framework that does UHR remote-sensing understanding as a visual perception problem through reasoning, reflection to diagnose insufficient evidence and guide the observation.

B) Authors construct a GeoRflex-48K which contains the trajectories to supervise the generation, teacher diagnoses supervise reflection, and restore the state to supervise the next correct trajectory using fallback tool which restores the parent view.

C) Author proposed recovery learned in the context when there error happens

D) The reflection reward penalizes the unnecessary reflections to avoid blind refection.

Critical weaknesses

A) The novelty is limited. The contribution overlaps heavily with prior work, for example zoom tools with GRPO are existing well established methods such as (Deepeyes, Geoeyes, GeoLens, GeoVista, ZoomEarth). However the parent-view fallback and judge reflection are minor contributions. But these can be seen as zoomout tool.

B) Authors claim where to look, ZoomEarth’s region-guided reward and GeoLens invoke the tools which tool to select and where to apply it.

C) I see that table 1 has different evaluation scores from GeoLens and ZoomEarth comparison tables against XLRS. Why it is so.

D) Table 2 does not compare scores only with format and accuracy. Since SFT and RL is performed, the gain with reflection and geometry reward might be minimum. Additionally, geometry rewards are common which are giving major gain whereas reflection reward is less contributing. Why it so?

Questions for authors

A) Why the scores are different in your table compared to GeoLens and ZoomEarth comparison tables against XLRS and similar trend is seen for all RSHR, XLRS, and MME-RS benchmarks,

B)- It is better to see how much percentage the reflection is called. How often reflection and fallback happen on the full benchmarks, and accuracy on questions with and without reflection calls to better understand the reflection reward benefits.

C)- Additionally to see the score with format and accuracy rewards only.

D)- Why authors used Thinking model instead of stronger Instruct model?

Reviewer_3ag9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GeoReflex targets wrong crops in active perception for UHR remote sensing. A single policy explores, diagnoses insufficient evidence, restores the parent view, and continues. The authors build GeoReflex-48K and post-train Qwen3-VL-8B-Thinking with SFT and GRPO, achieving strong results on RSHR-Bench and XLRS-Bench.

Critical strengths
  1. Recovering from wrong crops is a real problem, and making reflection an executable action (restoring the parent view) is a clear design.
  2. The protocols are detailed, the limitations are stated honestly, and performance remains stable across interaction budgets.
Critical weaknesses
  1. The self-correction claim is weakly supported. Removing the reflection reward changes little, reflection rarely triggers, and the benefit of diagnosis over neutral view restoration is marginal in one run and negative in another.
  2. Whether the model learns when to reflect is unverified. Only triggered states are analyzed, and the judge relies on the reference answer and later turns, without human validation.
  3. Same-backbone baselines are missing, such as SFT on generation data only and vanilla GRPO without the recovery tool.
  4. External comparisons are not controlled: GeoReflex uses a stronger backbone and largely reuses GeoVista's training data.
  5. The motivating problem is not quantified: how often failures actually stem from wrong crops is never measured.
Questions for authors
  1. How does GeoReflex compare with same-backbone SFT and vanilla GRPO under the same data and budget?
  2. On the full benchmarks, how often is reflection missed, unnecessary, or harmful?
  3. Can the ablations and mechanism analysis be repeated over multiple seeds?
  4. What fraction of base-model failures are caused by wrong crops?
Reviewer_P86oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GeoReflex, an active-perception framework for ultra-high-resolution remote-sensing understanding. The central idea is to make reflection actionable: after diagnosing insufficient visual evidence, the policy can restore a parent view, retain the rejected-path context, and redirect exploration. The authors construct GeoReflex-48K, containing 47,655 examples of successful exploration and paired reflection–refinement supervision, and train a single Qwen3-VL-8B policy through supervised fine-tuning and reflection-efficient GRPO. Evaluation covers three benchmarks, reward ablations, matched-prefix recovery controls, and interaction-budget comparisons. The final model achieves scores of 48.80 on RSHR-Bench and 57.69 on XLRS-Bench, exceeding the strongest compared external baselines by 5.40 and 7.04 points, respectively.

Critical strengths

1.A well-motivated connection between evidence diagnosis and executable visual recovery. GeoReflex addresses a concrete limitation of sequential cropping: an incorrect observation can leave subsequent reasoning without the necessary context. Restoring the parent view while retaining the rejected-path explanation provides a coherent way to reconsider the observation path. Integrating exploration, reflection, and refinement into one 8B policy also avoids requiring an external critic at inference. This makes the framework practically meaningful and gives the recovery supervision a clear operational purpose.

  1. Solid experimental evaluation. The method obtains the best RSHR and XLRS scores against about 25 baselines spanning closed-source, open-source, remote-sensing-specific, and active-perception models. The authors also candidly document the cohort scope, single-seed evaluation, and run-to-run variation (Appendix G). This establishes practical value for UHR remote-sensing understanding.
Critical weaknesses
  1. The paper’s premise is that current agents fail mainly because they cannot recover after a wrong crop. However, it gives no error analysis of existing agents (e.g., GeoVista, DeepEyes) showing how many failures come from a wrong crop that fallback could fix, and how many come from misrecognition or reasoning errors. Without this evidence, the motivation is weak.

  2. Limited novelty, and the closest prior work is missing. Keeping a failed prefix and appending a correction is already a common way to teach self-correction, as in Agent-R (Yuan et al., 2025) and Mulberry (Yao et al., NeurIPS 2025). Pixel Reasoner (arXiv:2505.15966) goes further: it trains on wrong visual operations followed by corrections, which is very close to the “wrong crop → recover” setting. The fallback action itself also closely resembles ZoomEye’s (EMNLP 2025) backtracking, which zooms out to the previous view when the needed evidence is missing. None of these works is cited or compared. The authors should make clear what is new beyond applying these ideas to remote sensing.

Questions for authors
  1. Train Qwen3-VL-8B-Thinking on the same corpus without the reflection/refinement examples, with the same GRPO budget and no fallback tool. How does it compare with GeoReflex?
  2. What is the final model’s performance when fallback is disabled at inference?
  3. On the full benchmarks, how often is fallback invoked? What is the accuracy of trajectories with and without fallback? After fallback, how often does the answer change from wrong to right versus right to wrong?

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

Related papers

Discussion (0)

Sign in to comment.