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

Paper · arXiv 2608.09819 · Published August 10, 2026
LLM Evaluations and Benchmarks

Macaron-V1 is an open agent-model family for experiential intelligence: learning from experience in real environments and continuing to learn after deployment. It is organized around two system goals. Adaptation is pursued through recursive improvement of versioned model-harness pairs, where experience from one configuration is evaluated under an external contract and used to construct its successor. Collaboration is pursued via the Mixture-of-LoRA (MoL) architecture that freezes a base model, composes specialist LoRA adapters, and selects one LoRA per user turn. The flagship Macaron-V1-Venti (748B) combines a 744B GLM-5.2 base with four LoRAs for chat, agent, coding, and GenUI; the Qwen3.6-35B-based Macaron-V1-Tall (50B) uses the same design for local deployment. This report presents Macaron-V1 as a co-designed system spanning architecture, algorithms, and infrastructure. The MoL architecture supports continual learning through extensible LoRA specialists. The algorithm combines Model-Harness Co-design and recursive self-improvement loop, including the UI4A component-native GenUI harness, a stateful action substrate, versioned Harness Context Protocol contract, and the agentic RL framework MindForge.

Introduction. As pre-trained models mature, an increasing share of advances in capabilities such as agentic tool use and coding has come from post-training (OpenAI, 2026; Anthropic, 2026; Moonshot AI, 2025; GLM-5 Team, 2026). The effectiveness of post-training, however, is closely coupled to the environment in which a model is trained, evaluated, and deployed (Yao, 2025; Lee et al., 2026b; Karten et al., 2026). The data a model learns from, the tools it can use, the users it serves, and the interactions it must support jointly shape the behavior required of a deployed agent (Harrington et al., 2026; Karten et al., 2026; Lee et al., 2026b). These conditions vary across deployments and continue to evolve over time, as new knowledge becomes available, new domains emerge, and interaction state accumulates across episodes (Li et al., 2025a; Harrington et al., 2026; Jin et al., 2026; Lei et al., 2026).

Discussion / Conclusion. Adaptation is not the property of a single model; it is the property of the loop the model sits in (Silver and Sutton, 2025; Yao, 2025). The RSI cycle (Section 3.2) generates harder tasks, audits trajectories under a versioned harness configuration, evaluates candidate HCPs, and only then uses selected trajectories to update weights. The design description concerns this loop; the released checkpoint is one snapshot and does not demonstrate improvement across generations. Collaboration is not a multi-agent framework layered on top of a model; it is an architectural affordance. Because the base is shared and adapters are portable, the registry can admit specialists trained by different teams or personalized for different users on the same runtime (Mind Lab, 2026b; Lu et al., 2026). The Proxy’s routing interface is the proposed interoperability contract; this release tests only the four shipped specialists. Neither bet is settled by Macaron-V1. The design is inspectable in the released harness (https://github. com/MindLab-Research/Mixture-of-LoRA-Harness) and Macaron-V1-Venti weights (https://huggingface.

Synthesis notes that discuss concepts related to this paper 7

Notes in this collection most semantically related to the paper, ranked by cosine similarity.

Lines of inquiry this paper opens 24

Research framings built by reading the notes related to this paper — the questions it feeds into.

How should we design LLM systems to maintain alignment and control? How can AI agents autonomously learn and transfer skills across tasks? Why does finetuning cause catastrophic forgetting of model capabilities? What drives capability and cost efficiency in agent systems? How does AI assistance affect human cognitive development and reasoning autonomy? Why do reward structures fail to shape long-term agent learning? What pretraining choices and baseline capability constrain reinforcement learning gains? How do self-generated feedback mechanisms enable effective model learning? How should agents balance memory condensation to optimize context efficiency? Do harness improvements transfer across model scales or memorize shortcuts? How do we evaluate AI systems when user perception misleads actual performance? Can model confidence signals reliably improve reasoning quality and calibration? Why can LLMs generate ideas better than they evaluate them? Can ensemble evaluation methods reduce bias more than single judges? How do evaluation biases undermine LLM quality assessment systems? Can single-axis benchmarks accurately predict agent deployment success? Does externalizing cognitive work and state improve agent reliability? What memory abstraction level best enables agent knowledge reuse? Can self-supervised signals enable process supervision without human annotation?