The paper investigates how to scale a Transformer-based vision‑language model trained from scratch on multimodal inputs under a fixed compute budget. It finds that minimal loss follows a predictable compute law while optimal model size and token count obey power‑law scaling, with language and multimodal objectives showing distinct allocation trends. The study also proposes a budget‑compensation rule that allows additional multimodal tokens to offset text‑efficiency penalties, and demonstrates that native multimodal pre‑training improves spatial reasoning and few‑shot learning.
arXiv:2608. 05000v1 Announce Type: cross Abstract: Vision offers a critical axis for advancing foundation models, driving a shift towards natively unified multimodal pretraining.
VisionWeave introduces elastic visual representation weaving, a native capability for multimodal large language models that learns where and at what granularity to encode visual information. The method combines a gated spatial pooler for coarse representations with a granularity router that allocates content‑adaptive token usage, trained end‑to‑end on large‑scale data. Experiments on Qwen3.5‑4B and Qwen3.8‑27B show that VisionWeave can save 43.0% of tokens while preserving 98.9% of performance across eight benchmarks, and delivers significant throughput gains and latency reductions when deployed on the SGLang serving engine.
arXiv:2605. 05225v3 Announce Type: replace-cross Abstract: Mixture-of-Experts Multimodal Large Language Models (MoE MLLMs) suffer from a significant efficiency bottleneck during Expert Parallelism (EP) inference due to the straggler effect.
arXiv:2607. 25527v1 Announce Type: cross Abstract: Unifying visual understanding and generation in one model holds immense promise, but remains challenging and expensive due to heavy compute and data demands and conflicts between the visual features needed for these two capabilities.
arXiv:2607.28640v2 Announce Type: replace Abstract: Multimodal large language models (MLLMs) should exhibit cross-modal consistency, generating consistent responses given semantically equivalent inpu...
VIVAS is a new Vision‑Language Model pre‑training framework that addresses the lack of fine‑grained visual perception in existing VLMs. It introduces a unified token space and a dense‑structural‑semantic vision tokenizer that expands the textual vocabulary with visual tokens, enabling vision‑language unified autoregressive supervision over both visual details and linguistic content. Trained on 12.4 T tokens, VIVAS achieves state‑of‑the‑art results on 7 tasks and 39 multimodal benchmarks.
The paper investigates image tokenizers as the visual language of unified multimodal models by creating a controlled autoregressive testbed that tracks task‑specific validation losses during multimodal continual pretraining across text, image, text‑to‑image, and image‑to‑text predictions. It shows that losses must be analyzed by task, that the loss–performance relationship varies with the token space, and that better reconstruction does not always lead to stronger downstream performance. The study also demonstrates how tokenizer design choices—such as discriminator use, semantic supervision, and vocabulary size—affect joint modeling and downstream results.
arXiv:2607. 26596v1 Announce Type: cross Abstract: Multimodal large language models (MLLMs) have demonstrated remarkable capabilities by integrating visual and textual understanding within a unified transformer architecture.
UVU is a vision-language unified autoregressive framework that integrates visual supervision directly into the pre-training stage of multimodal large language models. By using continuous visual encoding and a large-scale iterative hierarchical clustering algorithm to build a pixel-level visual codebook, UVU enables lossless representation of visual inputs and autoregressive generation of pixel-level image tokens alongside textual tokens. This approach synergizes pixel-level visual perception with semantic-level visual understanding, allowing models to internalize visual reconstruction capabilities and improve multimodal understanding performance.
arXiv:2505. 18227v4 Announce Type: replace-cross Abstract: In Transformer architectures, tokens\textemdash discrete units derived from raw data\textemdash are formed by segmenting inputs into fixed-length chunks.
arXiv:2609.16722v1 Announce Type: new Abstract: Scaling Multimodal Large Language Models (MLLMs) to long-form video understanding is bottlenecked by the explosion of visual tokens, which saturates co...
PRISM is a training‑free framework that efficiently selects visual instruction data for multimodal large language models by addressing the anisotropy in visual feature distributions, which causes a Global Semantic Drift. By implicitly re‑centering visual semantics, PRISM removes the influence of global background features, cutting data‑selection and model‑tuning time to 30% of conventional pipelines while improving performance across eight multimodal and three language benchmarks, achieving a 101.7% relative gain over baseline models.