The computational and memory overheads associated with expanding the context window of LLMs severely limit their scalability. A noteworthy solution is vision-text compression (VTC), exemplified by frameworks like DeepSeek-OCR and Glyph, which convert long texts into dense 2D visual representations, thereby achieving token compression ratios of 2x–10x. However, the impact of this high information density on the core long-context capabilities of vision-language models (VLMs) remains under-investigated. To address this gap, we introduce the first benchmark for VTC and systematically assess the performance of VLMs across three long-context understanding settings: VTC-Retrieval, VTC-Reasoning, and VTC-Memory. We further introduce VTCBench-Wild to evaluate robustness across diverse rendering configurations. We comprehensively evaluate leading open-weight and proprietary models on our benchmarks. The results indicate that several VLMs can retrieve information from rendered text in short-to-moderate contexts, but this ability degrades at longer lengths, and the degradation is especially pronounced on tasks requiring binding and integrating dispersed evidence. This study provides diagnostic evidence of VTC and serves as a foundation for designing more efficient and scalable VLMs.

Multimodal Reward Models are critical for aligning Multimodal Large Language Models with human preferences. However, existing reward models offer limited interpretability and struggle to deliver reliable preference judgments in audio-centric scenarios. To address this gap, we propose Audio-Reward-R1, a reasoning-based reward model that supports both audio understanding and audio generation preference tasks. Audio-Reward-R1 formulates reward modeling as a structured reasoning process, producing dimension-level analyses prior to the final preference judgment. To enable reliable reasoning supervision, we construct a high-quality reasoning-enhanced audio preference dataset with structured reasoning traces aligned with preference annotations. Audio-Reward-R1 is first trained via supervised fine-tuning as a cold-start on this reasoning-enhanced dataset, followed by reinforcement learning with verifiable rewards. Finally, we introduce Audio-Reward-Bench, a human-verified evaluation benchmark for evaluating audio reward models across both audio understanding and audio generation preference tasks. Extensive experimental results demonstrate the effectiveness and reliability of Audio-Reward-R1.