MIRROR: Leveraging Multi-View Complementarity to Enhance Multimodal Geometric Reasoning
Visual language models often exhibit inconsistent modality dependencies when performing multimodal reasoning. This paper introduces MIRROR, a novel reinforcement learning method that leverages these inconsistencies as a source of complementary information. The authors find that for the same geometric problem, models follow different reasoning paths and fail in different ways across text-only, image-only, and combined text-image views — revealing complementary reasoning signals. To exploit this, they construct ODA-Data, a high-quality paired dataset with three distinct views, and design a self-supervised reverse KL divergence optimization strategy. The approach uses the best-performing view as a teacher signal to guide training on other views, effectively harnessing cross-modal complementarity. Experiments show MIRROR significantly outperforms standard reinforcement learning methods across multiple geometric reasoning benchmarks, improving both single-modality accuracy and cross-modal behavioral consistency, establishing a new paradigm for multimodal reasoning model optimization.
Background and Context
While large language models have demonstrated remarkable proficiency in logical deduction, vision-language models continue to face significant hurdles when tackling visual reasoning tasks. This challenge is particularly acute in geometric problems that can be represented through text descriptions, visual diagrams, or a combination of both. The core research problem addressed in this study is the identification and resolution of behavioral inconsistencies exhibited by multimodal models across different input views. Through extensive observation, researchers discovered that models often display divergent reasoning performances when confronted with the same geometric problem presented in text-only, image-only, or combined text-image formats. For instance, a model might successfully solve a problem in a textual view but fail when presented with the corresponding image, and vice versa.
This inconsistency is not merely random noise but rather a signal indicating that different views expose complementary reasoning paths and specific failure modes within the model. Existing standard post-training methods for multimodal models have largely failed to leverage this phenomenon, resulting in bottlenecks in multimodal generalization capabilities. Consequently, this research makes a pivotal contribution by systematically quantifying these modality-dependent reasoning differences for the first time. It proposes a novel approach to self-improvement, aiming to enhance the overall performance of multimodal reasoning by mining complementary information across different views, thereby addressing the limitations imposed by ignoring modality-specific characteristics.
Deep Analysis
To address these challenges and effectively utilize the identified inconsistencies, the study introduces a dual innovation in both data construction and algorithmic design. First, the authors constructed ODA-Data, a high-quality paired multimodal geometric dataset specifically designed to study modality-dependent reasoning behaviors. This dataset contains three distinct views for each geometric problem: a text-dominant view, an image-dominant view, and a combined text-image view. It includes dedicated splits for training and evaluating modality-dependent behaviors, providing a robust foundation for analyzing how models process information across different modalities. This structured data allows for precise measurement of how reasoning shifts when the input modality changes.
Second, the study developed MIRROR, a Modality-Aware Inter-Reciprocal Reasoning Optimization method. MIRROR is a self-supervised optimization framework based on reinforcement learning. Its core mechanism revolves around "inter-reciprocity" and "teacher selection." For each problem, the algorithm evaluates the model's performance across all available views and dynamically selects the best-performing view as the "teacher" view. It then employs a reverse Kullback-Leibler (KL) divergence objective function to train the other, less performant views to mimic or align with the reasoning distribution of the teacher view. This strategy avoids the error propagation common in traditional methods, ensuring that the model always learns from the optimal solution for the current perspective. By doing so, it achieves mutual promotion and calibration of reasoning capabilities across different modalities without requiring external labels.
Industry Impact
The implications of MIRROR and the accompanying ODA-Data dataset extend significantly into both academic research and industrial applications. For the open-source community, ODA-Data fills a critical gap in high-quality multimodal geometric reasoning data. It provides a valuable benchmark for subsequent studies on modality dependence and reasoning consistency, enabling researchers to better understand and address the nuances of how vision-language models process complex information. This resource is essential for advancing the field of multimodal AI, offering a standardized way to evaluate and improve model robustness across different input types.
In industrial contexts, the relevance of MIRROR becomes even more pronounced as multimodal models are increasingly deployed in scenarios such as autonomous driving, robot navigation, and intelligent education. In these applications, consistency across different sensor inputs, such as camera images and text instructions, is paramount. MIRROR offers an effective pathway to enhance model robustness without incurring additional labeling costs. By leveraging the inherent complementarity between modalities, the method ensures that models behave reliably regardless of the input format. This capability is crucial for deploying AI systems in real-world environments where input conditions can vary widely, thereby reducing the risk of failures due to modality-specific blind spots.
Outlook
The success of MIRROR suggests a paradigm shift in how future multimodal models should be trained. Rather than focusing solely on aligning modalities, future research should delve deeper into exploring the complementarity and differences between them. By explicitly modeling and utilizing these differences, it is possible to construct more powerful and reliable general artificial intelligence systems. This approach moves multimodal reasoning beyond the basic capability of "seeing and speaking" toward a state of "logical rigor and consistency." The findings indicate that embracing inconsistency as a source of complementary information can lead to significant improvements in model performance.
Furthermore, the self-supervised nature of the MIRROR framework opens up new avenues for scalable training methods. As datasets grow in size and complexity, the ability to automatically identify and leverage the best-performing views for alignment becomes increasingly valuable. This could lead to more efficient training processes that require less human intervention and fewer external resources. The study also highlights the importance of dataset design, suggesting that future benchmarks should prioritize paired data with diverse modalities to better capture the nuances of multimodal reasoning. Ultimately, the integration of such techniques could pave the way for AI systems that are not only more accurate but also more robust and trustworthy in their decision-making processes across a wide range of applications.