This empirical research establishes the essential groundwork for predictably scaling multimodal foundation models by modeling the influence of data composition on compute laws and allocation exponents and derive an efficiency frontier specifying precise configurations of model size, token count, and data mixture.
Abstract
Although large language models (LLMs) exhibit remarkable reasoning capabilities, their reliance on text-only pre-training restricts the perception of the multimodal physical world. Native multimodal pre-training avoids this limitation by training models from scratch on multimodal inputs, thereby achieving deep cross-modal integration and mitigating optimization asymmetries inherent to traditional late-fusion architectures. Despite these advantages, the scaling properties of this paradigm remain systematically uncharacterized. To address this gap, we investigate the optimal model size and token count for training a transformer-based vision-language model under a fixed computational budget. We demonstrate that minimal objective loss adheres to a predictable compute law, whereas compute-optimal model sizes and token counts scale as power laws. Notably, language and multimodal objectives manifest distinct scaling behaviors. The language allocation law is largely invariant to the composition of the data, indicating stable language learning regardless of the multimodal data ratio. Conversely, the multimodal allocation law is highly sensitive to this composition. Specifically, text-heavy mixtures become compute-efficient only at larger model scales, shifting the optimal resource allocation toward greater model capacity. Additionally, by modeling the influence of data composition on compute laws and allocation exponents, we derive an efficiency frontier specifying precise configurations of model size, token count, and data mixture. Downstream evaluations further reveal that native multimodal pre-training induces positive cross-modal transfer, thereby enhancing pure-text spatial reasoning and enabling robust multimodal in-context learning. In summary, this empirical research establishes the essential groundwork for predictably scaling multimodal foundation models.
The results suggest that competitive MLLM can emerge from alignment alone, reducing multimodal extension to a lightweight projector-training problem that generalizes across modalities and adapts rapidly to each new LLM release.
Xuanru Zhou, Yiwen Shao, Jiahong Li et al.· 1 citation
SmartVL is proposed, a unified adaptive inference framework that jointly controls vision token number and model compute capability in response to varying input contents and compute budgets and consistently outperforms prior adaptive methods and achieves superior accuracy-efficiency Pareto frontiers.
Pengcheng Wang, Zhiquan Wang, Jayoung Lee et al.· 0 citations
Multimodal large language models have advanced rapidly, yet most remain English-centric, as scaling multilingual multimodal instruction tuning is limited by the scarcity and high cost of high-quality non-English image-text supervision. Although multilingual text data is abundant, naive textual fine-tuning can disrupt vision-language alignment and induce catastrophic forgetting. We propose Vision-Free Adaptation (VFA), a framework that decouples multilingual language enhancement from visual alignment by composing complementary task vectors over a shared LLM backbone. Specifically, we fine-tune a base LLM on multilingual text data to derive a multilingual task vector, which is then merged with the vision-aligned task vector of an MLLM. Experiments on five MLLMs across six multilingual multimodal benchmarks show consistent improvements while preserving both general multimodal and text-only capabilities. Moreover, using less than 2% of the text data, VFA narrows the gap to the fully multimodal-trained model, demonstrating its data efficiency.
Yixia Li, Yaqing Shi, Zhiwen Ruan et al.· Annual Meeting of the Associ...· 0 citations
While visual token pruning is essential for efficient Multimodal Large Language Models (MLLMs), existing training-free methods suffer from a critical limitation: they rely on static, instantaneous heuristics to perform irreversible filtering. This approach ignores the hierarchical nature of MLLMs, where token importance often evolves dynamically rather than remaining fixed across layers. Consequently, tokens essential for deep-layer reasoning are often prematurely discarded by shallow-layer estimates. To address this, we propose Trend-aware Pruning, a novel framework that elevates pruning from a local snapshot decision to a temporal trajectory modeling problem. Instead of relying on isolated scores, our method captures the momentum of attention flow. This enables a dynamic rectification mechanism that selectively reactivates"late-blooming"tokens, those initially undervalued but exhibiting rising semantic importance, thereby preventing the loss of critical visual cues. Extensive experiments demonstrate that our approach achieves a superior efficiency-performance trade-off across diverse multimodal tasks. Notably, it reduces visual tokens by over 77.8%, retaining only approximately 23 tokens in the final layer while maintaining competitive performance, offering a robust and reversible solution for high-efficiency multimodal inference.
The inference performance of Multimodal Large Language Models (MLLMs) typically relies on a massive number of visual block tokens, leading to substantial computational overhead. However, existing training-free compression methods often sacrifice semantic relevance, disrupt spatial coherence, or require extensive architectural modifications. Consequently, we propose Text-Conditioned Spectral Prototyping for Multimodal Token Compression (TCSP). We construct a joint affinity matrix based on both text relevance and visual similarity to derive boundary-continuous and semantically aligned segments via spectral embedding, subsequently replacing redundant information within these segments with weighted prototypes to reduce visual tokens for downstream inference. This method is entirely training-free and plug-and-play, allowing it to be inserted into intermediate layers of the vision encoder or before/after the projector module as needed. Evaluations on LLaVA-1.5-7B demonstrate that after a approximately $3 \times$ reduction in visual tokens, the overall performance remains baseline levels, even achieving marginal denoising-like improvements in several general QA benchmarks. More aggressive compression results in a graceful performance degradation, whereas tasks emphasizing compositional reasoning exhibit higher sensitivity to the compression scale. TCSP integrates text-awareness, spectral partitioning, and prototype replacement within a minimally invasive design. It achieves high, near-lossless compression ratios and a robust accuracy-efficiency trade-off without training, providing a concise and reliable pathway for general reasoning and resource-constrained deployment.