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

Recent progress in large-scale pre-trained models like GPT, T5, and GLM has demonstrated remarkable capabilities across NLP tasks. However, these models are predominantly static, trained offline on fixed datasets, and lack mechanisms for continual learning post-deployment. While techniques such as LoRA, MoE, and incremental fine-tuning have been proposed to improve adaptability, they often involve retraining or complex parameter updates, limiting scalability and efficiency. The need for models that can learn from real-world interactions, incorporate new knowledge dynamically, and collaborate across multiple specialized modules has become urgent. Prior efforts have addressed some aspects but lack a unified framework that supports lifelong, experiential learning in practical settings. This work builds on these foundations, aiming to create a scalable, modular system capable of ongoing self-improvement and collective intelligence.

The core challenge is enabling models to adapt continuously after deployment without retraining the entire network, which is computationally expensive and prone to catastrophic forgetting. Existing methods either update parameters incrementally, risking interference, or rely on external memory modules that are hard to scale. Additionally, integrating multiple specialized skills in a cohesive manner remains difficult, especially when aiming for real-time responsiveness and robustness in diverse environments. Achieving a system that can evaluate its own performance, identify areas for improvement, and incorporate new capabilities autonomously is a complex, multi-faceted problem that requires innovative architecture, algorithms, and infrastructure.

The key innovations include: 1) Mixture-of-LoRA (MoL) architecture, which allows the base model to remain frozen while dynamically composing specialized adapters, reducing retraining costs and enabling multi-task collaboration; 2) model-harness co-design, integrating environment-aware components like UI4A and versioned protocols for seamless interaction; 3) recursive self-improvement algorithms that evaluate and generate new model-adapter configurations based on external feedback; 4) infrastructure components like MinT and LongStraw, supporting scalable, long-context, and stable deployment. These innovations collectively enable a self-evolving AI capable of lifelong learning, continuous adaptation, and collective intelligence.

• �� Develop a large, frozen base model (e.g., GLM-5.2, Qwen3.6) as the core.
• �� Layer multiple specialized LoRA adapters targeting distinct tasks (chat, agent, coding, UI), each trained independently.
• �� Implement a Proxy-mediated routing loop where, at each user turn, the model classifies the request into one of the adapters based on a constrained decoding process.
• �� Use the entry adapter (L0) to classify requests, select the appropriate specialist, and generate responses within a three-stage cycle: route, answer, and summary.
• �� Integrate model-harness co-design, where environment feedback, user interactions, and versioning protocols guide iterative improvements.
• �� Employ recursive self-improvement: evaluate performance, generate new adapter configurations, and update the system based on external contracts.
• �� Infrastructure supports version control (MinT), long-context responses (LongStraw), and stability techniques for large sparse models.
• �� Validate system performance through benchmarks on personal intelligence, UI, and general tasks, comparing against state-of-the-art models.

Experiments involve benchmarking Macaron-V1 variants on datasets like Macaron ChatBench, LivingBench, and general capability suites. Metrics include accuracy, coherence, memory retention, and response quality. The models are tested across different configurations—varying LoRA ranks, adapter counts, and routing strategies. Ablation studies analyze the contribution of each component, especially the recursive self-improvement loop and infrastructure modules. Results show that recursive updates lead to cumulative performance gains, with over 15% improvement in multi-task benchmarks. The experiments also assess the efficiency of the routing mechanism, with accuracy exceeding 97% in adapter classification, and evaluate long-context reasoning capabilities, handling sequences over a million tokens.

The Macaron-V1-Venti model achieved a 15%+ cumulative improvement in multi-task benchmarks through recursive self-improvement, outperforming static models. The MiL architecture enables flexible, cost-effective multi-task collaboration, reducing fine-tuning costs by 80%. Long-context response evaluation demonstrated the model’s ability to handle sequences exceeding one million tokens with high coherence. The system’s ability to continually incorporate deployment experience resulted in steady performance gains over multiple iterations, validating the effectiveness of the recursive self-improvement mechanism. These results highlight the potential for scalable, self-evolving AI systems in real-world applications.

This framework is suited for personalized virtual assistants, enterprise AI solutions, and complex interactive UI systems. It enables continuous learning from user interactions, reducing the need for retraining. The modular design allows integration of diverse skills and knowledge bases, supporting multi-user and multi-task scenarios. Future applications include robotics, autonomous vehicles, and adaptive control systems, where ongoing self-improvement and environment adaptation are critical. The approach also facilitates collaborative AI development, where different teams can contribute specialized adapters, enhancing collective intelligence.

Despite promising results, the system faces challenges in robustness under adversarial or highly noisy conditions. Deployment costs remain high due to large model sizes and multiple adapters, limiting accessibility. The long-term stability of recursive self-improvement cycles is not fully validated, requiring further theoretical and empirical analysis. Additionally, the current infrastructure may not yet support real-time adaptation in resource-constrained environments. Future work must address these issues to enable broader, more reliable deployment in diverse real-world settings.