AIAny
Icon for item

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

A research report proposing a continual-learning agent workflow that pairs recursive self-improvement with a Mixture-of-LoRA design: freeze a foundation model, compose specialist LoRA adapters routed per user turn, and support them with long-context RL and post-training infrastructure.

Introduction

Most deployed agent models face a tradeoff between continually improving in the wild and avoiding destructive, monolithic model updates. This paper argues for a modular path: keep a strong frozen base model and add versioned, specialist LoRA adapters that can be post-trained, routed per user turn, and evolved by an agentic harness—so you get incremental capability gains without overwriting the foundation.

Key Findings
  • Mixture-of-LoRA (MoL) enables specialist composition: freezing a foundation model and loading per-turn LoRA adapters lets different experts (chat, agent, coding, GenUI) co-exist without catastrophic interference, meaning new capabilities can be added as plug-in LoRAs.
  • Recursive self-improvement via a model–harness co-design (MindForge + HCP) creates a closed loop where harnessed agents generate experience, evaluate successors under contracts, and post-train new versions—so improvement can be automated and versioned rather than done by ad hoc fine-tuning.
  • Infrastructure matters: MinT (post-training), LongStraw (multi-million-token RL training), and IndexCache/DSA stability techniques make long-context RL and sparse MoE training practical at scale; the paper validates this with flagship variants (Venti, Tall) and benchmark evaluations.
  • Routing and statefulness reduce context recomputation: L0 routing selects a specialist per turn and each LoRA keeps its own execution context, so continuity is preserved while avoiding repeated processing of full histories.
Who it's for and tradeoffs

Great fit if you are designing research or production agent systems that need continual capability extension without re-training a monolithic model, or if you plan to combine specialist behaviors (coding, UI actions, long-horizon agent tasks) on top of a shared foundation. Look elsewhere if you need a single, unified model that is continuously fine-tuned end-to-end, if you cannot afford the engineering complexity of a harness/LoRA deployment, or if you require formal guarantees about safety and distributional drift that go beyond empirical harness checks.

Where it fits

This work sits between traditional incremental fine-tuning and fully decentralized multi-agent stacks: it treats the base model as stable infrastructure and places the burden of change on small, composable adapters plus a versioned harness and evaluation contract. That makes it especially relevant for teams exploring modular continual learning, multi-specialist agent orchestration, and long-context RL at scale.

Information

  • Websitearxiv.org
  • OrganizationsMind Lab
  • AuthorsVin Bo, Asher Cai, Jingwei Cao, Song Cao, Vic Cao, Amelia Chen, Andrew Chen, Kaijie Chen, Cleon Cheng, Steven Chiang
  • Published date2026/08/10

More Items

Enables a coding agent to self-develop by evolving its harness, prompts, tools, and core code via reviewed commits — supporting recursive free evolution and experience-driven evolution. Demonstrated a 161-day live lineage and state-of-the-art scores on multiple coding benchmarks while foregrounding operational safety.

Describes a 314B-parameter decoder-only Mixture-of-Experts language model that activates 13.2B parameters per token for fine-grained sparsity, long-context (up to 256K) and multi-domain capabilities. Emphasizes GDLA architecture, expert balancing, and multi-teacher distillation.

Transfers hierarchical, structured memory from a large teacher agent into small LLM agents to improve tool-use success. Constructs Workflow, Subtask and Function memories with proactive/reactive injection; training-free and validated on multiple tool-use benchmarks.