AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Scaling Laws for Native Multimodal Models

Mustafa Shukor, Enrico Fini, Victor Guilherme Turrisi da Costa, Matthieu Cord, Joshua Susskind, Alaaeldin El-Nouby

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 41 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use early fusion instead of late fusionnative multimodal model 1.5B · SFT LLaVA mixtureAVG accuracy: 46.8 -> 47.6Improved
add mixture of experts (8 experts, top-1) to early fusion (ablation)early-fusion NMM 1.5B · SFT LLaVA mixtureAVG accuracy: 47.6 -> 48.2Improved
add MoEs to early fusion (ablation)early-fusion NMM 1.5B · SFT LLaVA mixture, COCOCIDEr: 72.0 -> 69.6Hurt
add MoEs to early fusion (ablation)early-fusion NMM 1.5B · SFT LLaVA mixture, TextVQAaccuracy: 28.1 -> 30.0Improved
freeze vision encoder instead of lr scaler 1 when initialized from pretrained (ablation)late-fusion NMM · SFT averageAVG (SFT) accuracy: 43.49 -> 54.14Improved
lower vision encoder lr scaler from 0.1 to 0.01 when training from scratch (ablation)late-fusion NMM · SFT averageAVG (SFT) accuracy: 34.84 -> 38.77Improved
use sparse MoE instead of dense for loss-compute fitearly-fusion NMM · pretraining validation lossmultiplicative constant: 29.574 -> 27.086Improved
use late fusion instead of early fusion for loss scaling exponentNMM · pretraining validation lossexponent c: -0.0492 -> -0.0494No clear effect
use sparsity (MoE) for loss scaling exponentearly-fusion NMM · pretraining validation lossexponent c: -0.049 -> -0.048No clear effect
use modality-agnostic MoE routing instead of modality-aware routing (ablation)sparse early-fusion NMM · image-caption and interleaved validation lossvalidation lossImproved
raise image-caption proportion in training mixture (30 to 40 percent) (ablation)early-fusion NMM · scaling exponent for image-captionabsolute exponent: 0.056 -> 0.061Improved
increase image resolution (ablation)early vs late fusion · pretraining lossvalidation lossMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy