Why it matters

Multimodal scaling shapes MLLM. Understanding shapes design.

Advertisement

The architecture

Modality mix in training.

Encoder + adapter + LLM staging.

Curriculum: text first, then multimodal.

Multimodal scalingModality mixtext + image + audioEncoder + adaptermodality bridgesCurriculumstaged trainingModality-specific scaling laws still being characterized; balance critical
Multimodal scaling.
Advertisement

How it works end to end

Encoder pretrained (CLIP, SigLIP, Whisper).

Adapter cheap; LLM optional freeze.

Curriculum + mixing.