acceptodds
Withdrawn from ICLR 2027

ReDiS: Reasoning–Diagnosis Self-Evolution for Multimodal Reasoning in Remote Sensing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal large language models (MLLMs) have shown strong potential for remote sensing understanding, yet reliable reasoning remains challenging when tasks require fine-grained visual grounding and multi-step inference. Existing approaches typically rely on either costly pre-constructed reasoning trajectories or self-generated signals such as response consistency, which can reinforce shared errors and provide unreliable supervision. We introduce Reasoning–Diagnosis Self-Evolution (ReDiS), a framework that separates reasoning exploration from candidate evaluation while allowing the two processes to evolve jointly. ReDiS instantiates two role-specific experts from a shared multimodal backbone: a Reasoner that explores diverse visual evidence–answer candidates and a Diagnoser that jointly evaluates the complete candidate set and provides credibility-aware feedback. As the Reasoner evolves, it exposes new candidate distributions and failure modes; the Diagnoser adapts to these changes and, in turn, reshapes the learning signal for subsequent exploration. To handle groups in which all sampled answers are incorrect and group-relative advantages vanish, we further introduce Failure-Triggered Recovery (FTR), which provides minimal outcome-level correction without prescribing intermediate reasoning trajectories. Experiments on LHRS-Bench, MME-RS, and CHOICE show consistent improvements over Qwen3.5 backbones at 0.8B, 2B, and 4B scales. ReDiS improves category-averaged accuracy by 6.47 percentage points at 0.8B, while ReDiS-4B achieves the highest average performance among the compared remote sensing MLLMs. Code and data will be released publicly.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote Sensing, Multimodal large language model, Self-Evolving

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

This paper introduces ReDiS, a post-training framework for remote sensing visual question answering. A Reasoner samples multiple evidence–answer candidates, and a Diagnoser evaluates those candidates and selects one or rejects the group. The two roles use separate LoRA adapters attached to a shared frozen Qwen3.5 backbone. The Reasoner is optimized with answer-correctness-gated rewards that include diagnostic feedback. The authors also propose Failure-Triggered Recovery (FTR): when an entire sampled group is incorrect and group-relative advantages provide no useful signal, answer-value tokens receive a supervised correction. Experiments on LHRS-Bench, MME-RS, and the multiple-choice subset of CHOICE show improved answer accuracy across 0.8B, 2B, and 4B models.

Critical strengths
  1. The motivation makes sense to me. In visual question answering, a model can give the right answer while using questionable visual evidence, and ordinary outcome rewards do not tell us where the reasoning went wrong. I can see why the authors introduce a separate Diagnoser and let its feedback influence training. Gating the reward by answer correctness is also a reasonable precaution.

  2. FTR is a useful part of the paper. When all sampled answers are wrong, the group-relative RL update has little to work with; giving a supervised correction on the answer tokens is a practical fix. The full method improves the 0.8B average score by about 6 points and also improves the 2B and 4B models. It is good that the extra diagnostic role is used during training and does not need to run at inference time.

Critical weaknesses
  1. I was not convinced that the Diagnoser has learned to judge visual evidence. In Section 3.3, the diagnostic targets are largely constructed from whether the final answer is correct. But a correct answer can still come with a made-up or irrelevant visual explanation. I may miss but I did not find an evaluation where evidence quality is checked independently of answer correctness. This is the part I would most like to see, because evidence diagnosis is central to the motivation. Also, It is difficult to tell how much of the improvement comes from the online Diagnoser.

  2. I am not sure why this method is presented specifically as a remote sensing approach. The paper mentions dense small objects and difficult spatial relations, but the Reasoner–Diagnoser loop and FTR seem applicable to general VQA almost unchanged. That is fine if the training method itself brings a sufficiently new result. In that case, though, the connection to earlier generator–verifier work such as RL Tango and hybrid RL/SFT training such as ReLIFT should be much clearer. If the contribution is primarily to remote sensing, I would like to know which of its harder question types benefit most. The category breakdown does not make that clear against the corresponding untuned Qwen3.5 models.

Questions for authors
  1. Could the authors evaluate the Diagnoser on a held-out set of responses with correct final answers but differing visual-evidence quality? Independent checks on a modest subset would be useful. I would substantially revise my view of the central contribution if the Diagnoser can consistently identify unsupported evidence under this control.

  2. The same training procedure appears applicable to general visual question answering. Which challenges of remote-sensing imagery are addressed especially well by ReDiS? A category-wise improvement analysis against each matched Qwen3.5 backbone, particularly for counting, localization, and spatial relations, could make this much clearer.

Reviewer_ycpmOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ReDiS, a reasoning–diagnosis framework for remote sensing VQA. A Reasoner and a Diagnoser, built on a shared Qwen3.5 backbone, are trained jointly: the Diagnoser's credibility scores refine the Reasoner's reward, while Failure-Triggered Recovery (FTR) supervises the answer tokens whenever all candidates are wrong.

Critical strengths
  1. Overall, the paper is clearly written and easy to follow, with sufficient implementation details for reproduction.
  2. FTR offers a simple and practical remedy for vanishing advantages in all-incorrect groups, a real issue in GRPO-style training.
Critical weaknesses
  1. My main concern is that the core claim does not hold. Since all rewards and targets are derived from reference answers, the method is standard RLVR rather than self-evolution.
  2. Novelty is also limited. Generator–verifier RL co-training already appears in RL Tango (NeurIPS 2025), and FTR is essentially a weaker version of ReLIFT (ICLR 2026); neither is cited. Moreover, nothing in the design is specific to remote sensing, and the paper does not demonstrate why this domain particularly needs such a framework.
  3. Regarding the experiments, the conclusions are not well supported. Every remote sensing baseline relies on a weaker backbone, the untuned Qwen3.5-4B already outperforms all of them, and none is fine-tuned on the same backbone and data.
  4. Ablations are insufficient as well. In all-correct groups the advantage depends solely on the noisy Diagnoser score, yet this case is never analyzed; moreover, removing the Diagnoser lowers avg OA by only 1.18.
  5. Finally, key baselines are missing, including SFT, vanilla GRPO/DAPO, and any self-evolution method.
Questions for authors
  1. Could the method be extended to genuine self-evolution that requires no answer labels?
  2. How does ReDiS compare with SFT, GRPO, and DAPO trained on the same backbone and data?
  3. GeoChat and VHM are built on LLaVA-1.5, which is far weaker than Qwen3.5. How can the gains over these baselines be attributed to the method rather than to the backbone?
  4. Two or three self-evolution methods (e.g., Co-rewarding, ICLR 2026) trained on the same data should also be included.
  5. The compared general-purpose models are somewhat outdated. Would ReDiS remain competitive against more recent open-source and closed-source models, such as Qwen3.8 and the GPT-6 series?
  6. What happens if the advantage of all-correct groups is set to zero?
Reviewer_vA3UOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReDiS, a two-role collaborative post-training framework for remote-sensing multimodal question answering. The Reasoner generates visual evidence–answer candidates online, while the Diagnoser evaluates the full candidate set and selects the most reliable response. Both roles share a frozen Qwen3.5 backbone and are optimized using separate LoRA adapters. The paper evaluates models at three scales (0.8B, 2B, and 4B) on LHRS-Bench, MME-RS, and the text-based multiple-choice subset of CHOICE, reporting accuracy improvements over the corresponding backbones.

Critical strengths
  1. The two-role self-evolution framework is novel and well motivated. ReDiS separates reasoning exploration from candidate diagnosis. The Reasoner explores diverse visual evidence–answer candidates online, while the Diagnoser jointly evaluates them and provides feedback to refine candidate preferences. Unlike approaches that rely on preconstructed reasoning trajectories or simple self-consistency signals, this design offers a promising direction for improving the reliability of remote-sensing multimodal reasoning.
  2. FTR addresses a key limitation of group-relative optimization. When all sampled answers are incorrect, the learning signal can vanish. FTR addresses this case through minimal outcome-level correction while preserving the freedom to explore intermediate reasoning. The ablation results and training dynamics suggest that FTR provides additional performance gains and helps stabilize training, demonstrating its practical value.
  3. The experimental results demonstrate effectiveness across benchmarks and model scales. ReDiS consistently improves over the corresponding Qwen3.5 backbones on LHRS-Bench, MME-RS, and CHOICE at the 0.8B, 2B, and 4B scales. The 0.8B model improves average AA by 6.47 percentage points, while ReDiS-4B achieves the highest average OA and AA among the compared models. Overall, these results support the effectiveness and potential of ReDiS for remote-sensing multimodal question answering.
Critical weaknesses

The main concerns fall into three areas:

  1. The aggregation protocol for evaluation metrics needs clarification. The category-level results for CHOICE appear to differ from the overall AA reported in the main table. Although this discrepancy may arise from differences in category hierarchies, weighting, or multi-label scoring, the current description is insufficient to reproduce the reported aggregate results. Further clarification would improve the transparency and verifiability of the evaluation.
  2. The implementation details and operational definitions of several training components are not sufficiently clear. The construction of FTR correction targets, token masking, loss normalization, and the exact configuration of the w/o Diagnoser ablation are difficult to reconstruct from the current description. The role of the router during training and inference also needs clarification. These issues mainly affect reproducibility and the attribution of gains to individual components.
  3. Some performance gains and training statistics lack sufficient quantitative characterization. Small OA/AA improvements are reported without corresponding paired uncertainty estimates. In addition, the relationship between the all-incorrect group rate, the proportion of correction-update steps, and actual training time needs to be clarified. These analyses would strengthen the empirical conclusions without requiring changes to the core method. The specific concerns are detailed in the questions below. If the authors can address them, I would consider increasing my rating.
Questions for authors
  1. Equation (7) defines the masked cross-entropy loss for FTR, but the actual correction procedure for the Reasoner and Diagnoser remains difficult to reconstruct. Could you provide a complete JSON correction target and a corresponding token-mask example for each role? Please clarify which tokens contribute to the loss, whether unsupervised evidence/reason fields remain as autoregressive conditioning inputs, and whether their content comes from the original sampled response or the correction target.
  2. Which training components are removed in the w/o Diagnoser ablation? Does it disable the Diagnoser's candidate generation, diagnosis, and own optimization entirely, or does it only remove the credibility feedback passed from the Diagnoser to the Reasoner? Clarifying this setup would help determine whether the observed performance differences can be attributed specifically to diagnostic feedback.
  3. The paper states that the router selects a role based on the input prompt, while the main evaluation activates only the Reasoner. What role does the router play during training? What are its input representation, architecture, training labels, loss function, and optimization procedure? Are the router and the two role-specific LoRA adapters trained jointly or independently?
  4. When the 21 categories displayed in Tables 5 and 6 are combined with equal weights, the resulting AA for ReDiS-4B is approximately 74.30%, whereas Table 1 reports 75.03%. Given the appendix's discussion of category-field fallbacks and multi-label scoring, does this discrepancy arise from differences in category mapping, weighting, or evaluation scope? Could you clarify the category hierarchy, sample counts, and aggregation formulas used in each table?
  5. Could you estimate 95% paired confidence intervals for the main OA/AA differences using the existing per-question predictions? Please clarify how predictions are matched by question ID, how correlations among multiple questions from the same image are handled, and how multi-label scoring and cross-benchmark averaging are incorporated.
  6. Appendix A.3 reports that 3,137 all-incorrect groups account for 11.53% of 27,200 source-question presentations, while correction updates occur in 823 of 850 Reasoner training steps. What does each statistic measure, and why do the all-incorrect group rate and the proportion of steps with correction updates differ substantially? Clarifying the denominators and timing boundaries would help distinguish the all-incorrect group rate, correction-update frequency, and actual training-time overhead.

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

Related papers

Discussion (0)

Sign in to comment.