Vibe Coding DiscoverAReaL
View on GitHubThe RL Bridge for LLM-based Agent Applications. Made Simple & Flexible.
AReaL is distributed infrastructure for asynchronous reinforcement learning of LLMs and agent applications. It includes training and inference services, supports agentic workflows and multiple RL algorithms, and integrates with SGLang, vLLM, and Ray.
Use Cases
Train LLMs for mathematical reasoningTrain multi-turn agents with reinforcement learningTrain coding agents on software engineering tasksRun online RL for black-box agent applicationsTrain search and customer-service agentsScale distributed training across GPU clusters
Built With
- Language
- Python
- Frameworks
- PyTorch · Transformers · SGLang · vLLM · Ray · LangChain · OpenAI Agents SDK · Qwen-Agent
Tags
reinforcement learning · agentic RL · LLM training · asynchronous training · distributed training · reasoning models · multi-turn agents · GPU