Awakening Visual Reasoning: Mitigating Post-Training Failure in Vision-Text Compression

Vision-Text Compression (VTC) offers a scalable path for long-context multimodal modeling by rendering textual data into dense visual tokens. While recent Vision-Language Models (VLMs) demonstrate high decoding fidelity (OCR) on such inputs, they exhibit a severe reasoning gap: models that reason robustly on native text often fail on visually compressed equivalents, particularly in long-range retrieval and multi-step deduction. We identify a phenomenon of post-training transfer failure, where standard supervised fine-tuning and reinforcement learning on visual prompts yield marginal gains compared to their textual counterparts. To address this, we propose CoRe (Coordinated Reasoning), a training framework that enforces lockstep consistency between the reasoning processes of textual and visual modalities. By treating the text-conditioned policy as a dynamic anchor, CoRe aligns the visual-conditioned policy via step-wise distribution matching, seamlessly integrating into both SFT and RL pipelines. Extensive evaluations across mathematical reasoning, long-context memory, and tabular retrieval benchmarks show that CoRe significantly outperforms standard visual post-training, recovering up to 70% of the performance gap relative to the textual upper bound and effectively activating latent reasoning capabilities in the compressed visual modality.

Modern multimodal LLMs can struggle with very long documents because processing long text is expensive and slow. One possible shortcut is to turn text into an image so the model can read more information at once. However, we found that this often creates a new problem: the model may still read the words in images correctly, but it does not reason over them as well as it does with normal text. We introduce a training method called CoRe to reduce this gap. When the same content is shown both as text and as an image, CoRe teaches the model to follow a similar step-by-step reasoning process in both forms. In this way, the stronger text version helps guide the image version during training. Our results show that this makes long-text understanding with multimodal LLMs more reliable on math problems, long documents, and tables. This could help build more efficient models that handle large amounts of information without losing as much reasoning ability.