PARL

# PARL: Parallel-Agent Reinforcement Learning

[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE)
[![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)
[![PyTorch](https://img.shields.io/badge/PyTorch-2.0+-ee4c2c.svg)](https://pytorch.org/)

> **⚠️ Disclaimer**: This is an **open-source community implementation** of the PARL (Parallel-Agent Reinforcement Learning) technique based on the Kimi K2.5 technical report. This is **NOT an official implementation** from Kimi AI or any affiliated organization. This project is maintained independently by The Swarm Corporation and the open-source community.

Open-source implementation of **PARL (Parallel-Agent Reinforcement Learning)**, a novel training paradigm that enables AI models to decompose complex tasks into parallel subtasks and coordinate multiple agents simultaneously.

## Overview

PARL is a training methodology that addresses the critical challenge of **serial collapse** in multi-agent systems, where models default to sequential execution despite having parallel computational capacity. By implementing staged reward shaping and a latency-oriented evaluation metric, PARL trains models to efficiently orchestrate up to 100 sub-agents across 1,500+ coordinated steps.

### Key Features

- **Staged Reward Shaping**: Dynamic reward annealing that encourages parallelism early in training and gradually shifts focus toward task success
- **Instantiation Reward**: Incentivizes subagent creation and concurrent execution
- **Critical Steps Metric**: Latency-oriented evaluation inspired by parallel computation's critical path concept
- **Differentiable Components**: Fully compatible with gradient-based optimization
- **Orchestrator-Subagent Architecture**: Trainable coordinator with frozen execution agents

## Architecture

```
┌─────────────────────────────────────────────┐
│         Orchestrator Agent                  │
│  (Trainable Central Coordin