Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA
This paper introduces Macaron-V1, an open agent-model family designed for continual learning and self-improvement in real environments, featuring a Mixture-of-LoRA architecture that freezes a base model while dynamically composing specialist adapters, alongside a recursive Model-Harness Co-design framework and supporting infrastructure for recursive adaptation and evaluation.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Learning Machine That Never Stops Growing
Imagine you have a brilliant new friend who knows everything about the world up to the day you met them. They are smart, but if you ask them about a new app released yesterday or a personal habit you developed last week, they are stuck. They can't learn from your specific conversations because their brain is "frozen" in time. This is the current limit of most artificial intelligence: they are trained once, shipped out, and then they just repeat what they know, unable to truly grow from their daily experiences.
The field of computer science trying to fix this is called Continual Learning. Think of it like trying to teach a student who never stops going to school. Instead of just memorizing a textbook and taking a final exam, a truly intelligent agent needs to learn from every interaction, every mistake, and every new tool it picks up, all while remembering who it is and what it's good at. The big question researchers are asking is: How do we build an AI that doesn't just "know" things, but actually experiences them and gets better at being a personal assistant over time?
This is where Macaron-V1 comes in. It's a new family of AI models designed to be "experiential," meaning they learn from real-life interactions and keep improving even after they are deployed. The researchers behind it, Mind Lab, realized that to make an AI truly smart, you can't just dump more data into a giant brain. You need a system that can adapt its environment and collaborate with specialized helpers. They built a system that treats the AI not as a single, static brain, but as a flexible team that can swap out tools and refine its own instructions in a loop of self-improvement.
The Macaron-V1 Team: A Frozen Brain and a Swappable Toolkit
Imagine a massive, super-smart library (the "base model") that never changes its books. It's so huge and well-organized that it knows almost everything. But, a library is just a library; it doesn't know how to help you specifically. To make it useful, Macaron-V1 adds a layer of "specialist adapters" on top of this library. Think of these adapters like different pairs of glasses or toolkits you can snap onto the library's head.
One pair of glasses is for chatting (L0), another for acting as an agent to do tasks (L1), a third for coding (L2), and a fourth for designing user interfaces (L3). This setup is called Mixture-of-LoRA (MoL). Instead of retraining the whole giant library every time you want it to learn a new skill, you just swap in a new pair of glasses or tweak the existing ones. The main library stays frozen and safe, while the specialists do the heavy lifting. This means the AI can have a "chat" personality one minute and a "coding" personality the next, without getting confused or forgetting its core knowledge.
The system uses a smart "Proxy" (like a receptionist) to decide which specialist to call. When you ask a question, the receptionist reads it, figures out if you need a coder, a designer, or just a chat buddy, and instantly switches to the right specialist. This happens so fast that you barely notice the switch. The paper shows that this system is incredibly accurate at picking the right specialist—about 99% of the time—and it doesn't slow things down too much, adding only a tiny bit of delay.
The Self-Improvement Loop: Tuning the Instructions, Not the Brain
The real magic of Macaron-V1 isn't just having these specialists; it's how the system optimizes its performance. The researchers built a system called MindForge that runs a "recursive self-improvement" loop. Imagine a video game where the AI plays a level, fails, and then the game automatically rewrites the rules or gives the AI a new strategy guide to try again.
Here is how it works:
- Discovery: The AI looks at what it can't do yet and creates harder challenges for itself.
- Expansion: It tries to solve these challenges in a simulated environment (like a sandbox). If it fails, the system doesn't just say "wrong." It analyzes why it failed. Was the instruction unclear? Was the tool missing?
- Update: If the AI finds a way to succeed by changing the instructions or the tools (the "harness"), it records that. If it needs to get better at the actual skill, it trains a new version of its specialist adapter.
The paper tested this on a set of 122 difficult tasks that the base AI couldn't solve at all. In a specific experiment described in the report, the model's brain remained completely frozen—no internal weights were changed. Instead, the system used configuration search to find the perfect combination of instructions, tools, and settings. By just tweaking these external instructions and tools, they managed to find a configuration that allowed the system to pass 100% of the tasks (122 out of 122).
It is crucial to note what this experiment proves: it demonstrates that the system can achieve full configuration-search coverage on these tasks. This means the AI already possessed the underlying capability to solve the problems, but it needed the right environment or prompt to unlock it. The experiment isolated this "Expansion" stage to show that changing the setup works, but it did not prove that the model learned the skill in a way that would generalize without those specific instructions, nor did it test the full self-improvement loop where the model's actual weights are updated.
The "Living" Benchmarks: Testing Real-World Smarts
To see if Macaron-V1 is actually good at being a personal assistant, the researchers created new tests called Macaron ChatBench and Macaron LivingBench. These aren't just multiple-choice quizzes.
- ChatBench tests if the AI can have a natural, empathetic conversation that feels like it's talking to a real person, not a robot. It checks if the AI understands your mood, stays honest, and moves the conversation forward.
- LivingBench is a simulation of real life. Imagine the AI is your personal assistant trying to plan a trip. The "world" changes around it: flights get delayed, prices go up, and you suddenly change your mind. The test sees if the AI can adapt its plan on the fly, verify information, and keep you calm.
In these tests, Macaron-V1-Venti (the big 744B version) scored very well, beating several other top-tier models in personal intelligence and coding. For example, it scored 87.8 on the UI4A-Bench (designing interfaces), while the next best was around 75.9. It also scored 94.0 on PinchBench, a test for complex tool use.
The Results: What We Know and What's Still a Mystery
The results are promising, but the paper is careful not to overhype them. The system works great at what it was tested on:
- Adaptation: It successfully found configurations to pass 100% of a difficult set of tasks by just changing instructions and tools (configuration search), proving that finding the right setup is a powerful way to improve performance without retraining the brain.
- Collaboration: The "Mixture-of-LoRA" approach works. The AI can switch between chat, coding, and design modes effectively without losing its mind.
- Efficiency: By keeping the main brain frozen and only loading small specialist tools, the system saves a massive amount of computer memory (about 74% less than running four separate giant models).
However, the paper also points out what it doesn't know yet.
- Long-term Learning: While the system can improve in a loop, the paper doesn't prove that it can keep doing this forever without getting worse at old tasks (a problem called "catastrophic forgetting"). They measured the potential for improvement via configuration search, but the long-term stability of continuous parameter updates is still an open question.
- Collective Intelligence: The system uses specialists trained by the same team. The idea of having specialists from different teams or users working together (like a chat bot from one company and a coding bot from another) is a future goal, not something tested here.
- Real-World Deployment: The tests were done in controlled simulations. We don't know yet how it will handle the chaos of the real internet or millions of different users with unique needs.
The Bottom Line
Macaron-V1 is a bold step toward AI that doesn't just "know" things but "experiences" them. By combining a frozen, stable brain with swappable specialist tools and a system that constantly rewrites its own instructions to solve problems, the researchers have built a model that is more adaptable than anything we've seen before. It's like giving an AI a Swiss Army knife and the ability to invent new tools on the fly.
The paper shows that this approach works: it solves hard tasks, designs better interfaces, and handles complex conversations better than many current leaders. But it also reminds us that this is just the beginning. The dream of an AI that learns from every single interaction, forever, without ever forgetting, is still a work in progress. The "Macaron" is baked, but the recipe for true, endless learning is still being perfected.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.