Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation

A new framework called ViGOS aims to prevent multimodal large language models from ignoring images in favor of text shortcuts during self-distillation training, according to a preprint posted on arXiv [1]. On-policy self-distillation (OPSD) trains a model on its own generated outputs, using a frozen copy of itself to provide dense token-level targets conditioned on a reference answer. The technique has proven effective for reasoning in large language models (LLMs) [2]. LLMs are machine learning models with many parameters, trained on vast amounts of text for natural language processing tasks [8]. However, researchers found that directly extending OPSD to multimodal large language models (MLLMs) introduces a vulnerability: the model can learn to rely on the text reference target rather than the image when generating tokens, creating a shortcut [2]. The ViGOS framework, detailed in a paper submitted to arXiv on June 17, 2026, decouples perception from reasoning to counter this shortcut [1]. The student model first writes a visual description of the image, then reasons toward a final answer. For valid rollouts, an image-only perception teacher supervises the description, while a privileged reasoning teacher supervises the reasoning and final answer using the same student-generated prefix. A reference teacher is employed only for invalid rollouts to recover the output format [2]. The authors evaluated ViGOS across five benchmark categories: general vision-language tasks, expert reasoning, visual math, spatial grounding, and visual-language-prior benchmarks. The framework retained the main benefits of OPSD while improving image-grounded behavior in settings prone to shortcuts [2]. The paper appeared on arXiv, an open-access repository of electronic preprints that is moderated but not peer-reviewed [6]. arXiv was founded in 1991 and now receives approximately 24,000 submissions per month as of November 2024 [6]. The platform also hosts arXivLabs, a framework for community-contributed tools that appear on article pages, though new proposals for arXivLabs are currently paused while the development team focuses on modernizing the site's infrastructure [3][4].