GitHub - Human-Agent-Society/reef: Continual learning infra for self-improving agents
Continual learning infra for self-improving agents - Human-Agent-Society/reef
Machine Learning Project
Summary
This repository presents a recursive reinforcement learning pipeline where a large AI agent is trained to design and execute training jobs for smaller models. The system utilizes a dual-loop RL architecture where the performance of the trained small model serves as the reward signal for the larger trainer agent.
Highlights
- Implements a dual-loop RL architecture (Outer loop for the trainer, Inner loop for the target model)
- Fully open-sourced, including LoRA weights, reward code, and GPU orchestration scripts
- Demonstrates successful skill transfer to held-out task families the agent had not previously encountered
- Leverages Tinker and GRPO for the reinforcement learning processes
auto-generated
Context
Audience
Machine Learning Researchers and AI Engineers
DomainMachine Learning
Formatopen source repository
Accessopen source
Source Overview
Open-source repo for an RL-trained agent (Qwen3.6-35B-A3B, LoRA) that writes complete prime-rl training jobs for smaller models on Runpod GPUs.
Trained via GRPO over 54 steps using Tinker, reward rose from ~0.0 to ~0.63 and transferred to a held-out task family. Shipped with agent LoRA weights, training scripts, six task families, reward code, and pilot write-ups.
Topics
Related
Reinforcement LearningGRPOQwen ModelsLoRAAutomated Machine LearningAgentic AI
