EARCP is a lightweight, powerful orchestration library for Large Language Models (LLMs) and complex decision-making systems. It enables you to create "virtual experts" from a single model or distinct modules, dynamically weighting their outputs based on consensus (coherence) and quality (performance).
Turn one model into a robust team of experts without fine-tuning.
## Core Concept: Single Model, Multiple Behaviors
**The Problem with Fine-Tuning**
Traditional approaches to specialize LLMs involve "Fine-Tuning": training a new model for every sub-task. This is expensive (GPU time), rigid (requires re-training for changes), and creates an MLOps nightmare (managing dozens of multi-GB files).
**The EARCP Solution**
EARCP extracts **diversity** from a **single, frozen model**. By modulating inference parameters like `Temperature`, `Top-P`, or `System Prompts`, we create "Virtual Experts":
* **Factual Expert ($T=0.1$)**: Deterministic, precise.
* **Creative Expert ($T=1.3$)**: Divergent, imaginative.
EARCP dynamically mixes these behaviors. **Result:** You get the benefits of an Ensemble Mixture-of-Experts (MoE) with simple software configuration, **zero training cost**, and minimal infrastructure.
## Industry Applications
EARCP is not limited to text generation. It is a general-purpose **Ensemble Decision Framework**. Here is how it applies to critical industries.
### 1. Robotics: Precision vs. Safety
In robotics, a manipulator arm must balance speed, precision, and safety. A single controller often struggles to optimize all three simultaneously.
**The EARCP Setup:**
* **Expert A (Speed):** Optimized for minimizing movement time (Aggressive).
* **Expert B (Precision):** Optimized for exact positioning (Slow).
* **Expert C (Safety):** Optimized for collision avoidance (Conservative).
**Orchestration Logic:**
EARCP continually evaluates the *Coherence* (do the trajectories agree?) and *Performance* (is the path valid?). during a delicate operation, the "Safety" and "Precision" experts will naturally gain higher weight, overriding the "Speed" expert.
```mermaid
graph LR
Input["Target Coordinates"] --> ExpSpeed["Speed Controller"]
Input --> ExpPrec["Precision Controller"]
Input --> ExpSafe["Safety Controller"]
ExpSpeed --> EARCP
ExpPrec --> EARCP
ExpSafe --> EARCP
EARCP -- "Dynamic Weighting" --> Output["Final Motor Command"]
style Output fill:#f9f,stroke:#333,stroke-width:2px
Self-driving cars must process perception data, legal constraints, and passenger comfort.
The EARCP Setup:
- Perception Expert: Identifies obstacles (Pedestrians, cars).
- Regulation Expert: Hard-coded traffic rules (Stop signs, speed limits).
- Comfort Expert: Smooth acceleration/braking curves.
Scenario: A car detects a sudden obstacle.
- Perception screams "STOP".
- Comfort suggests "Slow down gently".
- Regulation says "Stay in lane".
- EARCP detects low coherence but high performance risk from Perception. It shifts weight immediately to Perception + Regulation, executing a hard emergency stop, sacrificing Comfort.
graph TD
SensorData["Sensor Data"] --> Perception["Perception Expert"]
MapData["Map Data"] --> Regulation["Regulation Expert"]
IMUData["IMU Data"] --> Comfort["Comfort Expert"]
Perception -- "Obstacle Detected" --> EARCP["EARCP Core"]
Regulation -- "Legal Constraint" --> EARCP
Comfort -- "Smooth Trajectory" --> EARCP
EARCP -- "Conflict Resolution" --> Actuators["Vehicle Actuators"]
In medical diagnosis, avoiding "hallucinations" (false positives/negatives) is critical.
The EARCP Setup:
- Symptom Expert: Analyzes patient reports (NLP).
- Lab Expert: Analyzes blood work and numerical data (Statistics).
- History Expert: Checks patient's past records (Knowledge Graph).
Orchestration Logic: If the Symptom Expert suggests "Condition X" but the Lab Expert strongly contradicts it (Low Coherence), EARCP lowers the confidence score and flagging the case for human review, rather than outputting a confident but wrong diagnosis.
graph TD
PatientData --> SymptomExp["Symptom Analysis"]
PatientData --> LabExp["Lab Results Analysis"]
PatientData --> HistExp["History Cross-Check"]
SymptomExp --> EARCP
LabExp --> EARCP
HistExp --> EARCP
EARCP -- "Weighted Diagnosis" --> Doctorreport
For general chat and reasoning, we use the standard implementation to reduce hallucinations.
The EARCP Setup:
- Factual Expert (T=0.1): Ground truth.
- Creative Expert (T=1.2): Ideation.
- Critic Expert (System Prompt): "You are a critic. Verify inputs."
Experts are just wrappers around your model (LLM or otherwise).
import openai
from earcp import EARCP
class LLMExpert:
def __init__(self, temperature, system_prompt):
self.temperature = temperature
self.system_prompt = system_prompt
def predict(self, prompt: str) -> str:
# Example using OpenAI API
response = openai.chat.completions.create(
model="gpt-4",
messages=[
{"role": "system", "content": self.system_prompt},
{"role": "user", "content": prompt}
],
temperature=self.temperature
)
return response.choices[0].message.content
# Initialize a diverse team for General Purpose
experts = [
LLMExpert(0.1, "You are a factual assistant. Be precise."),
LLMExpert(1.2, "You are a creative assistant. Think outside the box.")
]earcp = EARCP(
experts=experts,
alpha_P=0.9, # Performance smoothing factor
alpha_C=0.8, # Coherence smoothing factor
beta=0.7, # Weight given to Performance vs Coherence
eta_s=4.0, # Softmax sensitivity
w_min=0.05 # Minimum weight per expert
)user_prompt = "Explain the impact of quantum computing on cryptography."
responses = [exp.predict(user_prompt) for exp in experts]
final_response, internal_weights = earcp.predict(responses)
print("Final Response:", final_response)
print("Current Expert Weights:", internal_weights)EARCP calculates a dynamic weight
-
Performance (
$P$ ): How good was this specific answer? (External signal) -
Coherence (
$C$ ): How much does this answer agree with the other experts? (Internal signal)
EARCP is not just a heuristic; it is grounded in Online Learning Theory.
Unlike traditional MoE (Mixture of Experts) which often requires experts to share architecture, EARCP is completely agnostic.
- Expert A could be GPT-4.
- Expert B could be a lightweight Llama-3-8B.
- Expert C could be a Random Forest classifier. EARCP orchestrates them solely based on their outputs and performance scores, enabling powerful heterogeneous ensembles.
We provide a theoretical guarantee on the Regret, defined as the difference between the cumulative loss of the EARCP algorithm and that of the best fixed expert in hindsight.
EARCP guarantees a sub-linear regret bound
This software is released under the Business Source License (BSL 1.1).
- Free for Non-Commercial Use: Research, personal projects, and testing.
- Commercial License Required: For production deployments.