Skip to content

Meta-Learning

Help small models punch above their weight

CRP includes three meta-learning mechanisms that improve model output without retraining. By decomposing hard tasks, injecting reasoning scaffolds, and reusing successful traces, CRP lets smaller, cheaper models compete with much larger ones - and helps every model improve over time.

Self-hosted today Managed-cloud waitlist for Gateway and Comply; more endpoints on the roadmap

The Problem

Small models (0.5B–4B parameters) struggle with complex tasks:

  • They forget instructions mid-generation
  • They produce shallow, generic outputs
  • They don't break problems into steps

CRP's meta-learning addresses this by providing external scaffolding that compensates for model limitations.

Three Mechanisms

graph TD
    A[Task Arrives] --> B{Model Capable?}
    B -->|Yes| C[Standard Dispatch]
    B -->|No| D[ORC: Break into Steps]
    D --> E[Execute Each Step]
    E --> F[Synthesize Result]

    A --> G[ICML: Build Envelope]
    G --> H{Trace Library?}
    H -->|Yes| I[Add Few-Shot Examples]
    H -->|No| J[Add Reasoning Scaffold]
    I --> C
    J --> C

    C --> K[Result]
    K --> L{Quality ≥ 0.7?}
    L -->|Yes| M[RTL: Store Trace]
    L -->|No| N[Discard]

ORC: Orchestrated Reasoning Chains

ORC decomposes complex tasks into micro-steps when the model can't handle them in one shot.

When ORC Activates

All three conditions must be true:

Condition Check
Resource pressure Must be < HIGH (won't add overhead under pressure)
Model capability < task complexity Model needs help
Probe quality < 0.7 Initial attempt was poor

If the model can handle it alone, ORC stays out of the way.

How ORC Works

  1. Decompose - Split the task into 3–10 micro-steps based on task type
  2. Execute - Dispatch each step individually (each gets its own envelope)
  3. Synthesize - Combine step outputs into a coherent final answer
  4. Retry - If a step fails, retry with heavier scaffolding

Scaffold Levels

Level Name What It Does
0 None No scaffolding
1 Light Key factors to consider
2 Medium Step-by-step template
3 Heavy 5-step structural template with explicit output format

Example

A 2B model asked to "Compare React, Vue, and Angular":

Without ORC

Produces a shallow 3-paragraph summary missing key dimensions.

With ORC

  1. Step 1: "List the core architecture of React" → focused output
  2. Step 2: "List the core architecture of Vue" → focused output
  3. Step 3: "List the core architecture of Angular" → focused output
  4. Step 4: "Compare these architectures on performance, learning curve, ecosystem" → structured comparison
  5. Synthesis: Combine into coherent comparison article

Each step gets the model's full attention on a manageable subtask.

ICML: In-Context Meta-Learning

ICML enriches the dispatch envelope with reasoning scaffolds and few-shot examples from the trace library.

Adaptive Scaffolding

The scaffold level adapts to model capability:

Model Capability Scaffold What's Added
≤ 1 (0.5B–1B) Heavy 5-step template: "1. Identify key concepts, 2. List relevant facts, 3. Organize by theme, 4. Draft response, 5. Review for completeness"
≤ 2 (2B–7B) Light Key factors prompt: "Consider these factors: ..."
> 2 (7B+) None Model doesn't need scaffolding

Few-Shot Examples

If the trace library has similar past tasks with quality ≥ 0.7:

  1. Retrieve top-3 matching traces (by word overlap)
  2. Include their step descriptions as examples in the envelope
  3. Model sees "here's how a similar task was solved before"

Metacognitive Envelope

The final envelope combines:

Base Envelope (facts, task, system prompt)
  + Reasoning Scaffold (if model needs it)
  + Few-Shot Traces (if available)
  = Metacognitive Envelope

RTL: Reasoning Template Library

RTL stores successful reasoning traces for future reuse.

Storage Criteria

A trace is stored only if:

  • Quality score ≥ 0.7 (configurable via rtl_min_quality_for_storage)
  • The trace is complete (all steps executed)

Trace Structure

ReasoningTrace(
    trace_id="abc-123",
    task_type="comparison",
    task_summary="Compare React, Vue, Angular",
    steps=[
        ReasoningStep(
            step_description="Analyze React architecture",
            system_prompt_template="...",
            expected_output_format="bullet_points",
            scaffold_level=1
        ),
        # ...
    ],
    model_class="2B-7B",
    quality_score=0.82,
    usage_count=0
)

Retrieval

When a new task arrives, RTL retrieves traces by word overlap:

  1. Tokenize task description
  2. Compare against stored trace summaries
  3. Return top-k matches (default: 3)
  4. Increment usage_count on retrieved traces

Curation

Every curation_interval (default: 5) requests, the library prunes:

  • Traces with quality below threshold
  • Traces that haven't been used

From the SDK

import crp

client = crp.SDKClient()

# Choose depth to control how much reasoning scaffolding is applied.
client.configure(depth="thorough")

# For weak local models, use a stricter meta-learning profile.
answer = client.ask("Compare React, Vue, and Angular", depth="thorough")
print(answer.quality)
print(answer.how_it_was_built)   # STL / ORC operation sequence
print(answer.open_questions)

Advanced Configuration

from crp.advanced.meta_learning import MetaLearningEngine, MetaLearningConfig

config = MetaLearningConfig(
    enabled=True,
    orc_enabled=True,
    orc_max_steps=10,
    orc_min_model_capability=1,
    icml_enabled=True,
    icml_max_examples=3,
    rtl_enabled=True,
    rtl_min_quality_for_storage=0.7,
    scaffold_level="auto",       # "auto"|"none"|"light"|"heavy"
    curation_interval=5
)

engine = MetaLearningEngine(
    dispatch_fn=my_dispatch_function,
    model_capability=2,           # 0=tiny, 1=small, 2=medium, 3=large
    config=config
)

Model Capability Levels

Level Parameter Range Examples
0 < 0.5B TinyLlama
1 0.5B – 1B Phi-2, Qwen-0.5B
2 2B – 7B Qwen3-4B, Llama-3.1-8B
3 7B+ Llama-3.1-70B, GPT-4

Research Foundations

Meta-learning in CRP is grounded in:

  • MAML (Finn 2017) - Learning to learn with few examples
  • ICL as Implicit Gradient Descent (Dai 2023) - In-context learning IS learning
  • STaR (Zelikman 2022) - Self-taught reasoning via bootstrapping
  • Distilling Step-by-Step (Hsieh 2023) - 770M model outperforming 540B with scaffolding

The key insight: you don't need a bigger model if you give a smaller model better structure. CRP's meta-learning provides that structure automatically.

See Research Foundations for the full academic basis.