acceptodds
Withdrawn from ICLR 2027

MIRROR: Learning from the Other View for Multi-Modal Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language models (VLMs) can behave very differently depending on whether the same underlying problem is presented textually or visually. This inconsistency suggests that different views expose different reasoning paths and failure modes that standard multimodal post-training does not fully exploit. To study and exploit this phenomenon, we construct ODA-Data, a high-quality paired multimodal geometry dataset with text-dominant, image-dominant, and combined image+text views of the same problems, together with splits for training and evaluating view-dependent reasoning behaviors. We then developModality-Informed Reciprocal Reasoning Optimization (MIRROR), a reinforcement learning approach that derives a cross-view self-distillation signal from the model’s own modality-dependent capabilities. For each problem, MIRROR evaluates the model under all views, selects the best-performing view as a teacher, and trains the remaining views using a GRPO task-reward objective augmented with a reverse-KL regularizer toward that teacher. Across reasoning benchmarks that evaluate on multimodal mathematics problems, MIRROR improves over standard RL and yields more accurate and consistent behavior across different input modality representations.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Mathematics Reasoning, Reinforcement Learning, Vision Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_e89XOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper shows that VLMs often solve the same geometry problem in one view but fail in another. It builds a paired-view geometry dataset and proposes MIRROR. MIRROR picks, for each problem, the view the model currently solves best as a teacher, and trains the weaker views with GRPO plus a distillation term toward that teacher. On Qwen3-VL and GLM-4.1V models, it reports gains over single-view and mixed-view GRPO on ODA-Val, GeoInt, and MathVerse.

Critical strengths
  1. A good testbed and a clean method. Showing the same problem in different views gives a controlled way to study cross-modal failures, and the paper shows that no single view is always best. The student always generates its own answers, and the teacher view only scores them, which avoids copying information the student cannot see. A slowly updated teacher is shown to be necessary for stable training.
  2. Sound experimental practice. The experiments use current models from several families, add a second domain (charts), and report compute costs honestly. They also include a useful negative result: simply mixing views in GRPO does not transfer reasoning between them.
Critical weaknesses

1. The closest prior work is not cited, which narrows the contribution. arXiv 2510.22836 (Oct 2025) already distills from the model's own full-text view into its image-dependent view inside RL, with the same loss weight (0.01), on geometry and MathVerse. This is similar to MIRROR with a fixed text teacher. TIGER (ICML 2026) already selects problems that are solved as text but failed as images and trains on the image version with GRPO, which closely matches how ODA-Data is filtered. Neither is cited or compared. This is consequential because what remains new is choosing the teacher view per problem and the slowly updated teacher, so the paper's value depends on showing that these matter.

2. The remaining contribution, choosing the teacher per problem, is not established. Adaptive selection beats the best fixed teacher by only about 2 points on ODA-Val, and it loses on MathVerse and on GeoInt pass@16. All results are single runs, and the distillation weight was chosen on the test sets. About half of the time there is no teacher at all, and the image view acts as teacher only about 5% of the time, so the "reciprocal" transfer is mostly from views that contain the text. The headline consistency gain is measured on the training problems. This is consequential because, without evidence that per-problem selection beats a fixed teacher, the paper's distinctive claim is unsupported.

Questions for authors
  1. How does MIRROR compare with the recipe of arXiv 2510.22836 and with TIGER, under identical data and backbone?
  2. Can you report multi-seed results for the adaptive vs. fixed-teacher ablation (Table 3), with the distillation weight chosen without the test sets? Please also give the teacher distribution per student view, and both-view solvability on held-out data.
Reviewer_1F2eOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper identifies view-dependent reasoning asymmetry in VLMs and proposes MIRROR, which dynamically selects the strongest view of each problem as a teacher to guide weaker text- or image-dominant views through on-policy reverse-KL distillation combined with GRPO. The authors also construct ODA-Data, a paired multimodal geometry dataset with aligned text, image, and combined views. Experiments on ODA-Val, ODA-Unfiltered, GeoInt, MathVerse, larger VLMs, and ChartX show that MIRROR generally improves reasoning accuracy and cross-view consistency over standard and mixed-modality GRPO baselines.

Critical strengths
  1. Clear and useful problem formulation. The paper turns cross-view inconsistency from a robustness issue into a training signal, using the strongest view of each problem as an online teacher for weaker views. This is a simple but meaningful idea that is well matched to the observed modality asymmetry.

  2. Strong and reasonably broad empirical validation. MIRROR consistently outperforms standard and mixed-modality GRPO on the main geometry benchmarks, and the gains further generalize to larger VLMs and to chart reasoning, suggesting that the benefit is not limited to a single model or dataset.

Critical weaknesses
  1. The contribution of cross-view transfer is not fully isolated from additional regularization and computation. MIRROR introduces not only adaptive cross-view teacher selection, but also EMA-based reverse-KL distillation and about 37.5% additional FLOPs over mixed-modality GRPO. The paper also lacks a same-view EMA distillation baseline, making it difficult to determine how much of the improvement is specifically attributable to cross-view supervision.

  2. The assumption of strong and generalizable complementarity across modality views is not sufficiently justified. Prior visual-math studies, including MathVista [1], MathVerse [2], MathFlow [3], and CogFlow [4], generally show that MLLMs rely more reliably on textual information, while visual inputs often introduce perception and internalization errors. MIRROR itself also reports substantially stronger average performance for text-dominant than image-dominant inputs. More importantly, ODA-Data explicitly retains examples where one view fails while another succeeds, thereby enriching the dataset for exactly the asymmetry MIRROR is designed to exploit. This raises concerns that the method may overfit to specially mined cross-view failure cases and generalize less effectively to natural visual-math distributions.

  3. The online teacher-selection mechanism may be noisy. MIRROR selects the strongest view using only 16 rollouts per view and empirical success rates, so small performance differences may reflect sampling variance rather than a genuine capability advantage. In addition, teacher selection is based on the current policy, while the actual guidance is produced by an EMA model. The paper does not sufficiently analyze rollout-count sensitivity or the consistency between these two rankings.

[1] Lu, P., Bansal, H., Xia, T., et al. MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts. ICLR, 2024.

[2] Zhang, R., Jiang, D., Zhang, Y., et al. MathVerse: Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems? ECCV, 2024.

[3] Chen, S., Yuan, H., Xu, Y., et al. MathFlow: Enhancing the Perceptual Flow of MLLMs for Visual Mathematical Problems. ACL, 2026.

[4] Chen, S., Xu, Y., Xie, J., et al. CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving. ICLR, 2026.

Questions for authors
  1. Where do the gains from cross-view supervision actually come from?
    Could the authors provide a same-view EMA distillation baseline and a strictly compute-matched GRPO baseline? Since MIRROR simultaneously introduces cross-view teacher selection, EMA-based reverse-KL distillation, and additional computation, these controls are important for isolating the contribution of cross-view supervision itself.

  2. How prevalent is the cross-view asymmetry exploited by MIRROR in natural, unfiltered distributions?
    Since ODA-Data is explicitly constructed by retaining cases where one view succeeds while another fails, it would be important to report how frequently such complementary behavior occurs without this filtering, and whether MIRROR still provides clear gains on such data. This would help assess whether the method generalizes beyond specially mined cross-view failure cases.

  3. How stable is the online teacher-selection mechanism?
    How sensitive is teacher selection to sampling noise with K=16K=16 rollouts per view? Does the selected teacher remain stable when varying the rollout count, and how often does the view ranking under the current policy agree with that under the EMA teacher actually used for distillation?

Reviewer_tR4pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MIRROR to improve vision-language reasoning when the same problem is presented through different input views. It constructs ODA-Data with paired text-dominant and image-dominant geometry problems, with a third combined view. During training, the model attempts each view, selects the best-performing view, and uses an EMA copy conditioned on that view to guide student-generated responses alongside GRPO. The proposed contribution is to turn differences between views into a problem-specific training signal without a separate external distillation teacher. Validation includes ODA-Val, unfiltered ODA problems, GeoInt, MathVerse, larger base models, and a separate chart-domain training experiment.

Critical strengths
  1. The paper addresses a meaningful and practical problem with a clear methodology. A model should not fail to solve a problem simply because the input format changes from text to an image.The authors introduce a straightforward training mechanism, that they use the model's performance differences between text and image inputs to let the stronger modality guide the weaker one. 
  2. The experiments are comprehensive and demonstrate clear performance gains. Compared to the strongest standard GRPO baseline, the proposed method increases the image-view final-answer correctness (pass@1) from 20.38% to 23.57%, and the text-view from 41.06% to 45.67%. Furthermore, the experimental workload is substantial. They successfully validated their method on Qwen3-VL-8B and GLM-4.1V-9B models, expanded the evaluation to chart-domain tasks, and investigated different teacher models and training settings.
Critical weaknesses
  1. A central claim of this paper is that the proposed method improves not only the overall accuracy but also the model's consistency and robustness across different representations. However, this claim is entirely unverified on test sets consisting of held-out problems that did not participate in the training phase. In Section 5 (Result 3, p. 8), the paper claims that the proportion of problems solved under both views increased from 42.5% to 60.7%. However, this metric is evaluated solely on the training set (ODA-Train). Furthermore, when evaluating the held-out datasets (ODA-Val and ODA-Unfiltered) in Table 2 (p. 7), the paper only reports the text-view accuracy and image-view accuracy separately. An increase in individual scores cannot directly prove the claim that "if the model can solve a specific problem in one view, it can still reliably solve the exact same problem when the input is switched to another view."
Questions for authors

Q1. If text-based input question lacks necessary conditions, the model's failure to solve it should be attributed to a flawed prompt, not a lack of reasoning ability. If the image-based version yields better results, it may simply be because the image inadvertently supplies the missing context. In appendix K.5. The text states that the squares "share a common side." Geometrically, this typically implies sharing a complete, equal-length edge. However, the provided image displays squares of different sizes with only partial boundary overlap. Furthermore, the provided condition (Area B + Area C = 400) is mathematically insufficient to uniquely determine that the sum of their perimeters is 112. The image does not clearly depict any additional geometric constraints required to uniquely derive this answer.

Q2. The paper provides contradictory descriptions of how the cross-view guidance algorithm resolves ties. Section 4 (p. 5): Pick the view with the highest success rate. If multiple views have the same score, pick one randomly. Appendix A (p. 16): If no other view has a strictly higher score than the current one, the algorithm triggers a "No teacher" condition. Algorithm 1 (p. 21): Cross-view guidance is skipped only if the algorithm randomly picks the current view itself as the teacher. For example: Three different views are each evaluated 16 times, and all yield 0 correct answers. According to the rule in Section 4, the algorithm would randomly assign one of these equally failing views to act as a teacher. However, according to Appendix A, the mechanism would assign "No teacher" because no alternative view is strictly better. These two rules trigger fundamentally different training behaviors.

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

Related papers

Discussion (0)

Sign in to comment.