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

Summary (Overview)

• Macaron-V1 is an open agent-model family designed for experiential intelligence—learning from real-environment experience and continuing to learn after deployment, rather than approximating a static optimum through centralized post-training.
• The system is built around two complementary goals: adaptation (recursive improvement of versioned model–harness pairs) and collaboration (via the Mixture-of-LoRA architecture that freezes a base model and composes specialist LoRA adapters).
• The flagship Macaron-V1-Venti combines a 744B GLM-5.2 base with four LoRA specialists (Chat, Agent, Coding, GenUI), while Macaron-V1-Tall (50B) uses the same design on a Qwen3.6 base for local deployment.
• Key innovations include the Model–Harness Co-design algorithm (UI4A component-native GenUI harness, REPL agent harness, Harness Context Protocol) and the MindForge agentic RL framework running a three-stage recursive self-improvement loop.
• Evaluation spans Personal Intelligence benchmarks (Macaron ChatBench, Macaron LivingBench), GenUI (UI4A-Bench), and general capability benchmarks, with Macaron-V1-Venti achieving the highest scores on TerminalBench 2.1 (87.6) and UI4A-Bench (87.8) among compared models.

Introduction and Theoretical Foundation

The paper motivates Macaron-V1 by observing that post-training effectiveness is closely coupled to the environment in which a model is trained, evaluated, and deployed. As pre-trained models mature, an increasing share of capability advances comes from post-training, but the resulting systems remain tied to the task and environment distribution available during training.

The authors define experiential intelligence as "the ability to learn from experience accumulated in a real environment and to keep learning after deployment." They operationalize this through two dimensions:

1. Adaptation: Recursive improvement over explicitly versioned model–harness pairs, where experience from one configuration is evaluated under an external contract and used to construct its successor. This contrasts with:

   • Parametric continual-learning methods (updating weights across tasks)
   • Memory/state-based methods (carrying trajectories forward without modifying parameters)
   • Context/harness-based methods (updating instructions, tools, and orchestration)

2. Collaboration: Pursued through composition via Mixture-of-LoRA (MoL)—a frozen base model with specialist LoRA adapters selected per user turn. This occupies a distinct operating point between mixture-of-adapter systems and multi-agent systems.

The theoretical foundation rests on the observation that joint post-training over heterogeneous tasks can create cross-task interference through shared parameters. The MoL architecture makes specialization explicit through a single design rule: "Cluster tasks that share skills and thinking patterns into one LoRA, and keep tasks whose skills diverge sharply in separate LoRAs."

Methodology

Mixture-of-LoRA (MoL) Architecture

Design: A frozen 744B GLM-5.2 base with four LoRA specialists, each with rank 16 and LoRA alpha 32:

• L0 (Chat): Conversational backbone, instruction following, and routing entry point
• L1 (Agent): Long-horizon, heavy tool-use tasks
• L2 (Coding): Code generation and terminal use
• L3 (GenUI): UI4A rendering and UI-driven action

Routing Loop: A three-stage lifecycle per user turn:

1. Route: L0 classifies the request into one of four canonical adapter labels under a 24-token decode budget with constrained decoding
2. Answer: The chosen specialist responds from its own conversation view
3. Summary: The specialist emits a ≤192-token summary stored server-side as shared context

Model–Harness Co-design

The harness is treated as a first-class optimization target:

• UI4A: Component-native GenUI harness where the agent writes ordinary frontend code (imports, components, state, Actions) under runtime-enforced boundaries. Actions have four fields: Origin, State, Execution, and Visibility (including NoAI boundaries).
• REPL Agent Harness: A stateful read–eval–print loop with persistent Python namespace, supporting executable composition and validated reuse (save_tool/promote_tool with validation).
• Harness Context Protocol (HCP): A versioned TOML contract standardizing runtime selection, action surface, context resources, session state, and environment contract.

Recursive Self-Improvement (RSI) Cycle

The agent policy is formalized as:

$$\pi_{\phi}(a_{t} \mid o_{\leq t}; \theta, c),\tag{1}$$

where $\theta$ = frozen base parameters, $\phi$ = trainable LoRA parameters, $c$ = versioned harness configuration, and episodes recorded as:

$$\tau = (o_{0}, a_{0}, \ldots, o_{T}, a_{T}, y),\tag{2}$$

The three-stage cycle:

1. Discovery: Model proposes harder task variants (lifting constraints, introducing hidden preferences, chaining sub-goals)
2. Expansion: Tasks executed under fixed model–HCP pair; audits localize failures to model, task, or harness
3. Update: Selected trajectories update LoRA specialists; accepted HCP registered as next runtime configuration

Infrastructure

• MinT: Post-training platform managing adapter revisions (immutable LoRA snapshots) and policy records (mutable service state)
• LongStraw: Response-only long-context execution with live-memory boundary:

$$M_{\mathrm{live}} \approx M_{\mathrm{weights}} + M_{\mathrm{prefix}}(P) + M_{\mathrm{grad}} + \max_{i} M_{\mathrm{graph}}(R_{i}) + M_{\mathrm{score}}\left(\sum_{i} R_{i}\right),\tag{3}$$

• Sparse-base controls: R3 rollout routing replay, DSA implementation alignment, and IcePop-style residual filtering

Empirical Validation / Results

Routing Performance

Table 1: Per-hop latency of the route–answer–summary loop (48 multi-turn requests, temperature 0)

Routing accuracy: 99.12% on a 6,448-sample trace for Venti; 99.04% for Tall. Residual errors concentrate at the L0/L1 boundary (general chat vs. personal-agent).

Expansion Coverage on Base-Failure Set

Table 4: Phase breakdown of the 69-job expansion run (122 tasks the frozen base fails)

Adaptive search over harness configurations reaches 122/122 coverage, while the best single-configuration sweep reaches only 11/122.

Main Benchmark Results

Table 8: Macaron-V1-Venti vs. six comparison models (0–100 scale; ∗ = imported public value)

Multimodal retention (Table 10): The text-only LoRA adapters on Macaron-V1-Tall show mixed results vs. the base—higher on OCRBench (+0.80), MMBench-EN (+0.60), MMMU (+1.33), and MME cognition (+66.43), but lower on MME perception (−52.99).

Case Study: Executable Composition

A discrete function-calling arm requires 48 turns for an order/ticket/SLA task, while the REPL completes the same task in 6 turns by retaining intermediate values in a persistent namespace.

Theoretical and Practical Implications

Theoretical contributions:

• Establishes MoL as an architectural substrate for continual learning and collective intelligence, decoupling release cadence of base, specialists, and harness
• Formalizes the distinction between language-space configuration search and parameter-space adapter training as complementary update paths
• Demonstrates that "failure under the baseline configuration does not by itself mean the behavior is absent from the frozen model"—configuration search can elicit latent capabilities

Practical implications:

• MoL reduces stored parameter residency by ~74% vs. replicated-base layouts (774.8B vs. 2.976T logical parameters for Venti)
• Routing overhead is stable (~32% of three-hop total) across base sizes
• The REPL substrate reduces turn counts by up to 8× on compositional tasks
• UI4A reduces output tokens by ~45% vs. raw HTML (672 vs. 1,224 tokens on average)

Key limitations acknowledged:

• Cross-generation compounding gains not yet demonstrated
• No controlled ablation attributing gains to specific components
• REPL performs worse on stateful observe-before-commit APIs (49.5% vs. 54.0% on BFCL v4)
• Qualitative character-stability degradation in very long LivingBench sessions

Conclusion

Macaron-V1 represents a deliberate architectural bet: orchestration over merging. The system treats continual learning not as a property of a single model but as a property of the loop the model sits in, with three release clocks (base, specialists, harness) moving independently.

The authors frame the central open questions for future work:

1. Scaling recursive self-iteration: Whether RSI gains compound across generations or converge to local modes
2. Emergence of collective intelligence: Whether independently trained adapter populations produce capability beyond any constituent specialist
3. Scenario richness: Closing the evaluation–deployment gap through more diverse scenarios and quantified stability analysis

The roadmap includes more specialists (domain-specific research, non-English long-form), more UI4A rendering targets (Flutter, native mobile), third-party adapter composition paths, and deeper integration with the artifact loop for real interaction data.

The paper concludes: "Building a growable agent model is not the same problem as building a strong general model"—requiring environment, harness, training loop, and model architecture to evolve as separately versioned but jointly evaluated layers. The current results validate the system's execution checks and one model snapshot, while longitudinal continual-learning gains and beneficial cross-population composition remain open empirical questions.