Macaron-V1: Continual Learning with Self-Improvement and Mixture-of-LoRA Adapters

Macaron-V1 represents a novel approach to continual learning in AI systems. Unlike frozen models that are trained, evaluated, and deployed as static products, Macaron-V1 treats deployed models as the starting point for an ongoing learning process. This architectural shift enables systems to learn and adapt in real-world environments, improving themselves through billions of interactions post-deployment.

Key to this approach is the concept of versioning and the development of a feedback loop that directly integrates into the system. By treating deployment as the beginning of a learning cycle, rather than the end, Macaron-V1 enables continuous improvement. Each version learns from production feedback, gets evaluated against a defined quality contract, and either advances to the next iteration or is discarded.

This cycle relies on the Model-Harness Co-design, which treats the model and its supporting environment as a single, versioned entity. The harness encompasses how users interact with the model, the tools it can utilize, how outputs are evaluated, and what constitutes success. By co-designing these components, Macaron-V1 ensures that improvements to the model directly enhance the system's user-facing behavior.

A central mechanism in this process is the evaluation contract—a formal, versioned specification of what constitutes "better" performance. This contract prevents drift and ensures that learning aligns with the system's underlying intentions. Versioning throughout the process allows for rollback capabilities, comparison of different approaches, and a clear record of improvements.

To add new capabilities without retraining the entire model, Macaron-V1 employs the Mixture-of-LoRA technique. This method allows for the addition of specialized parameters to the base model through low-rank adaptation, effectively enabling the system to grow in capability while retaining its foundational knowledge. By freezing a stable base and composing lightweight adapters, Macaron-V1 achieves significant improvements without sacrificing the system's existing knowledge.