Successive Capacity Growth: Task-Complexity-Driven Width and Depth Expansion for Vision Transformer Encoders in JEPA World Models

arXiv cs.AI · Frederik Berenz · 2026-08-27

This is no longer on the current shelf — shelves rotate as new material clears the bar. The analysis below is unchanged. See what is featured now.

Successive Capacity Growth (SCG) turns Vision Transformers into dynamic learners that grow their brains—width and depth—as tasks get harder.

What it does

SCG is a method for Vision Transformers (ViTs) in Joint-Embedding Predictive Architectures (JEPAs) that starts small (1 head, 2 layers, 283K params) and grows incrementally in width (adding attention heads) or depth (adding layers) as task complexity increases. The approach reduces redundancy, improves efficiency, and adapts model capacity to the problem at hand.

How it applies

Use SCG in edge devices, embedded systems, or anytime you need a model that adapts its capacity to the task. It’s particularly useful for JEPA-based world models in robotics or autonomous systems.

Takeaways

• SCG starts with minimal ViT encoders and grows capacity dynamically.
• Reduces redundancy in attention heads and layers.
• Adapts model size to task complexity, improving efficiency.
• Works with JEPA world models for adaptive, resource-aware AI.
• Promises lower compute costs and faster inference in dynamic environments.

The short version

Your model just grew a bigger brain—on the fly.

Why it matters

We’re drowning in over-provisioned models. Teams throw 10B-parameter beasts at problems that could be solved with 1B, just because they’re afraid of underfitting. SCG flips the script: start small, grow smart. It’s a blueprint for AI that learns to allocate compute like a human brain does—efficient, adaptive, and resource-aware. This is how we build agents that don’t melt GPUs.

My take

Teams waste millions on static, over-sized models that barely scrape by. SCG is the antidote. It’s not just about efficiency—it’s about teaching models to learn smarter, not harder. If you’re building agents for the real world, where compute is always a bottleneck, this is a game-changer. And it’s a reminder that we’re still in the Stone Age of model adaptability.

How it connects

• Aligns with the push for dynamic, resource-aware AI.
• Supports the rise of JEPA and world models in robotics.
• Highlights the inefficiency of static, over-provisioned models.

Bottom line

Stop over-provisioning your models—start small, grow smart with Successive Capacity Growth.

Topics: JEPA, ViT, adaptive-compute, world-models, efficiency