In the quest to build "lifelong" AI, researchers face a fundamental tension: a model must be flexible enough to learn new skills without being so volatile that it overwrites old ones. This challenge, known as catastrophic forgetting, has traditionally been met with two extremes. On one hand, we can fine-tune a single set of weights, which inevitably erodes previous knowledge. On the other, we can instantiate a new "expert" module for every new task.
While the modular approach—using Parameter-Efficient Fine-Tuning (PEFT) like LoRA or Adapters—effectively prevents forgetting, it introduces a new bottleneck: parameter bloat. If a model gains a new set of weights for every task, its storage requirements grow linearly. As Weixiang Zhao and colleagues note in SAPT: A Shared Attention Framework for Parameter-Efficient Continual Learning of Large Language Models, this linear accumulation becomes unsustainable as the sequence of tasks scales into the thousands.
The emerging research direction focuses on breaking this linear dependency, moving toward architectures that achieve sub-linear or even constant parameter growth while maintaining "zero-forgetting" performance.
The Linear Growth Trap
The intuition behind modular continual learning is simple: if you don't touch the weights used for Task A while learning Task B, Task A cannot be forgotten. This is the philosophy behind GAN Memory with No Forgetting by Yulai Cong et al., which freezes a base model and adds task-specific "style" parameters.
However, as Martin Wistuba and the team at AWS argue in Continual Learning with Low Rank Adaptation, even "efficient" modules like LoRA add up. If each task requires only 1% of the original model's parameters, after 100 tasks, you have doubled your model size. For an agent learning an open-ended stream of environments, this is a memory death sentence.
Strategy 1: Factorization and Shared Bases
One way to slow the growth is to realize that tasks are rarely entirely unique. If Task 10 is similar to Task 2, they should share parameters.
Xiaobing Yu et al. propose FM-LoRA: Factorized Low-Rank Meta-Prompting for Continual Learning, which factorizes adaptation weights to recognize task similarities. Similarly, CL-LoRA by Jiangpeng He and Fengqing Zhu 0001 introduces a dual-adapter architecture. It splits the learning into a task-shared component (capturing general features) and a task-specific component (capturing nuances). By maximizing the work done by the shared component, the system reduces the footprint of the new modules needed for each step.
Strategy 2: Dynamic Routing and Reuse
Instead of just adding modules, why not build a library of skills and learn how to combine them? This shifts the problem from storage to selection.
In Incrementally Learning Multiple Diverse Data Domains via Multi-Source Dynamic Expansion Model, Runguo Wu and colleagues use a dynamic graph-based router. When a new task arrives, the router attempts to reuse existing "experts" from previous tasks, only expanding the model when the new domain is sufficiently novel. This allows the parameter count to grow only when the model actually encounters new information, rather than simply new task labels.
Strategy 3: Consolidation and Merging
The most aggressive approach to the scaling problem is to "squash" the modular experts back into a single set of weights periodically. This is known as Model Consolidation.
Iordanis Fostiropoulos and Jiaye Zhu developed Batch Model Consolidation (BMC), which trains experts in parallel and then distills their knowledge into a single base model. This keeps the inference-time model size constant.
However, simple weight averaging often leads to "interference" where the features of one task degrade another. To solve this, Zecheng Pan introduced Merging without Forgetting: Continual Fusion of Task-Specific Models via Optimal Transport. By using Optimal Transport to align the "geometry" of different models before merging them, the framework can integrate new knowledge into a fixed-size backbone without the typical performance drop.
Taking this to the limit, Fuli Qiao and Mehrdad Mahdavi proposed SLAO, which uses a single shared LoRA adapter for all tasks. By using asymmetric merging and orthogonal initialization, they achieve the performance of a multi-adapter system with the memory footprint of a single one.
The Intuition: From Library to Synthesis
The shift in this field is a move from a "Library" metaphor to a "Synthesis" metaphor.
In the Library stage (early modular CL), we stored every task as a separate book on a shelf. It was safe, but the shelf eventually ran out of space. In the Synthesis stage (current research), we are learning to weave new threads into an existing tapestry. We use shared attention mechanisms to decide which parts of the tapestry to use (SAPT), and we use mathematical alignment to ensure that adding a new color doesn't ruin the existing pattern (OTMF).
The goal is a system where the cost of learning the 1,001st task is significantly lower than the cost of learning the first, as the model leverages a vast, consolidated foundation of previously acquired expertise.