Improving Multimodal Reasoning via Worst Dimension Optimization

A new preprint proposes a method called Worst Dimension Optimization to address a structural weakness in how multimodal AI systems learn to reason, aiming to prevent strong performance on one factor from masking failures on others [1]. The work, posted on the arXiv preprint repository, targets the design of Process Reward Models (PRMs) used in multimodal reasoning—tasks that require a system to maintain integrity across constraints such as visual grounding and logical consistency [1][2]. Current PRMs rely on heuristically defined rewards that weigh these factors equally, a practice the authors argue can allow dominating dimensions to conceal individual failures, leaving the overall validity of the reasoning process unguaranteed [2]. The proposed Worst Dimension Optimization approach shifts the focus to the dimension performing most poorly at any step, rather than averaging performance across all constraints [2]. The preprint does not include experimental results in its abstract, but frames the work as a conceptual reorientation for building more reliable reasoning paths in systems that process both images and text [1][2]. Multimodal models are a growing subset of foundation models—large-scale systems trained on vast datasets so they can be applied across a wide range of use cases [4]. Building such models is often highly resource-intensive, with the most advanced costing hundreds of millions of dollars to cover data acquisition, curation, and the compute power required for training [4]. The reasoning capabilities of these systems are frequently evaluated through benchmarks that measure factual accuracy, alignment, and safety [11]. The preprint appears within a broader research effort to improve the trustworthiness of AI reasoning. In many fields of mathematics and physics, almost all scientific papers are self-archived on arXiv before publication in a peer-reviewed journal [9]. As of November 2024, the repository was receiving about 24,000 articles per month [9]. The paper is accessible through arXiv’s abstract page, which includes community-developed tools under the arXivLabs framework—a formalized program that allows third-party collaborators to build experimental features such as citation explorers and code finders directly into the site [7][8]. arXivLabs was launched to enable innovative collaborations while ensuring partners share arXiv’s values of openness, community, excellence, and user data privacy [8]. The framework is currently on hiatus for new proposals while the development team focuses on modernizing arXiv’s infrastructure and moving systems to the cloud [6].