Infini-AI-Lab

Infini-AI-Lab / astraflow

Public

Dataflow-Oriented Reinforcement Learning for (Multi-)Agentic LLMs

infini-ai-lab.github.ioastraflowdocs
29
1
89% credibility
Found May 20, 2026 at 29 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

AstraFlow is a research system for training AI assistants to become better at complex tasks. It uses reinforcement learning - a type of training where AI learns by practicing, receiving feedback, and improving over time. The system supports training on various tasks including math problems, code writing, household chores in virtual environments, web shopping, and multi-agent collaboration where one AI checks another's work. It can run multiple training strategies simultaneously and scale across different computers.

How It Works

1
🔬 You discover AstraFlow

You learn about a system that can train AI assistants to become better at solving complex problems through practice and feedback.

2
📦 You install the system

You set up AstraFlow on your computer, which gives you all the tools needed to train AI models.

3
🎯 You choose what to train

You pick a training recipe - maybe math reasoning, code writing, or a multi-agent team where one AI checks another's work.

4
⚡ Training begins automatically

The system runs multiple AI models through practice problems, learning from their successes and failures in parallel.

5
🔄 Multiple policies collaborate

Different AI strategies train together, each improving by watching and learning from the others.

6
You test your trained AI
🏠
ALFWorld tasks

Your AI learns to navigate a virtual home, picking up and placing objects.

🛒
WebShop tasks

Your AI learns to browse and purchase items from an online store.

🎮
Game playing

Your AI learns strategic social deduction games with other players.

🎉 Your AI is ready

You now have a trained AI assistant that can solve complex problems, write code, or help with various tasks.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 29 to 29 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is astraflow?

AstraFlow is a dataflow-oriented reinforcement learning system for training agentic LLMs. It lets you run multi-policy collaborative RL training across distributed, heterogeneous hardware that spans multiple regions. The system handles the messy infrastructure work: you define your training recipe (math reasoning, code generation, web agents), point it at your models, and AstraFlow manages the async rollout collection, trainer coordination, and elastic scaling. It supports SGLang and vLLM as rollout engines, with Megatron-core for training backends.

Why is it gaining traction?

The hook is elastic cross-region rollouts. Unlike systems that assume fixed cluster topology, AstraFlow lets compute nodes join and leave the rollout pool on demand without scheduler-specific code. For teams running training across cloud regions or mixed hardware, this removes a significant operational headache. The composable data algorithms also stand out: you can mix GRESO, dynamic sampling, and buffer replay without rewriting training loops.

Who should use this?

ML engineers building multi-agent LLM systems who need distributed RL training at scale. Research teams working on RLVR or M2PO-style training for math and code tasks. Teams running across heterogeneous infrastructure or multiple cloud regions who want a unified training interface. Not yet suitable for teams needing production stability or offline cluster training (both on the roadmap).

Verdict

AstraFlow solves real scaling problems for LLM RL training and has a published paper backing its design. However, with only 29 stars and a credibility score of 0.8999999761581421%, it is extremely early-stage. The documentation is solid and the architecture is well-thought-out, but test coverage and production hardening remain unproven. Worth evaluating for research use cases, but wait for the all-in-one launcher and offline cluster support before considering production deployment.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.