New Method Scales AI Models Through Expert Blocks Without Retraining

A team of computer vision specialists has developed a technology that scales already-trained AI models into larger versions by adding specialized expert modules, without launching training from scratch.

What is the core idea behind the new method?

The new technology breaks the usual link between a model's power and the need to train it from scratch. The traditional approach goes like this: if you need a larger neural network, you design a new architecture, assemble a dataset, and run a full training cycle that can take weeks or months of compute. The team's method takes an existing pretrained model and expands it by adding specialized expert modules on top of the knowledge it has already accumulated.

The idea rests on keeping what already works. A pretrained model has already seen billions of data examples and has formed valuable internal representations of the world. Throwing those away for the sake of scaling is wasteful. The new method uses these weights as a stable foundation and builds specialized components on top.

The approach is conceptually close to the Mixture of Experts (MoE) architecture, in which different "experts" inside the network specialize in different types of data or tasks. The key difference is that the new solution grows such a structure out of a ready-made base rather than building an MoE from scratch.

"The technology makes it possible to expand pretrained AI models into larger systems with specialized expert modules without training from scratch," the authors say of their work, according to

TechXplore.

Why does this matter for AI developers?

The cost of training frontier neural networks is one of the industry's main barriers. According to analysts at Epoch AI, the cost of training leading models in 2024–2025 reached tens of millions of dollars. Teams that already have a well-trained base model face a difficult choice: work with a limited version, or spend a comparable sum on retraining from scratch.

Scaling through expert modules offers a third path. Existing models can be developed incrementally: new specialized blocks are added as tasks and requirements grow, without losing the accumulated weights and internal representations of the base network.

• The technology was created by a team of computer vision specialists
• It expands pretrained AI models without full retraining from scratch
• It uses specialized expert modules built on top of a ready-made base
• It preserves the weights and accumulated representations of the original model during scaling

Where does it apply?

Computer vision tasks are especially sensitive to model scale. Medical image recognition, industrial quality control, and autonomous driving systems all require both a broad general knowledge base and deep specialization in a specific domain. Expanding through expert modules addresses this trade-off: take a powerful general model and build specialized blocks on top of it for a particular task.

For enterprise teams that already have models trained on proprietary data, this means they can upgrade a system without losing company-specific knowledge and without going through a full cycle of data collection and training again.

What this means

If the expansion method with expert modules proves effective in practice, it could change how enterprise AI systems evolve: instead of costly restarts every few years, incremental growth in which each model version becomes the foundation for the next.