Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation
About
On-policy self-distillation (OPSD) trains a model on its own rollouts and uses a frozen copy to provide dense token-level targets conditioned on a reference target. This works well for LLM reasoning, but a direct extension to multimodal large language models (MLLMs) can create a shortcut: the privileged target may guide tokens mainly based on the text reference target rather than the image. We propose ViGOS, a visually grounded OPSD framework for MLLM post-training. The student first writes a visual description and then reasons toward the final answer. For valid rollouts, an image-only perception teacher supervises the description, while a privileged reasoning teacher supervises the reasoning and final answer on the same student prefix. A reference teacher is used only for invalid rollouts to recover the output format. Across general vision-language, expert reasoning, visual math, spatial grounding, and visual-language-prior benchmarks, ViGOS keeps the main benefits of OPSD and improves image-grounded behavior in shortcut-prone settings.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Expert/Academic Reasoning | MMMU | Pass@5 Score80.11 | 8 | |
| Expert/Academic Reasoning | MMMU-Pro | Pass@564.81 | 8 | |
| Prior-Sensitive Evaluation | ViLP-F with-fact | Score70.17 | 8 | |
| Prior-Sensitive Evaluation | ViLP-P pure-question | Score69.5 | 8 | |
| Spatial & Vision Grounding | MMSI | Pass@566.4 | 8 | |
| Spatial & Vision Grounding | CV-Bench | Pass@591.51 | 8 | |
| Visual Math | MathVista | Pass@580.9 | 8 | |
| Visual Math | MathVerse | Pass@568.91 | 8 | |
| General Visual Language | MM-Vet | Pass@572.02 | 8 | |
| Spatial & Vision Grounding | RealworldQA | Pass@586.8 | 8 |