rllm-org
rLLM is an open-source initiative from UC Berkeley's Sky Computing Lab that builds infrastructure to train, evaluate, and evolve intelligent agents with reinforcement learning. Its goal is to democratize RL for language models and agents by supporting arbitrary agent harnesses, environments, benchmarks, algorithms, and training backends.
At a Glance
- Academic AI/ML research labs
- Open-source model and agent researchers
- AI infrastructure and agent engineering teams
- Organizations training coding, research, finance, and other tool-using agents
- +1 more
AI Tools by rllm-org
(1)
rLLM
RL Training Framework for Agents
Discussions
No discussions yet
Be the first to start a discussion about rllm-org
Latest News
Continual Learning via Real-Time RL for Agents
Training a Frontier Java Code Migration Agent with AWS AgentCore Runtime
Hive: Collaborative Agent Evolution Platform
rLLM UI: Real-Time Observability for Agent Training & Evaluation
Products & Services
Apache-2.0 open-source framework and command-line workflow for evaluating and training language agents with RL. It supports arbitrary agent programs, 60+ integrated benchmarks, reward/evaluator functions, and the same code path for evaluation and training.
SDK/gateway layer that works with API-based agent programs and frameworks, captures exact token IDs and log probabilities, and translates calls into training-ready trajectories without requiring an agent rewrite.
Standardized post-training orchestrator spanning multiple agent abstractions, algorithms, and backends, including RL, SFT, and on-policy distillation with pluggable backend protocols.
Cloud-hosted or self-hostable web interface for real-time monitoring and analysis of rLLM training/evaluation runs, including episode and trajectory search, logs, code/configuration inspection, and an observability AI agent.
Market Position
rLLM positions itself as an open-source orchestration and training layer for agentic RL rather than a single model or a single inference/training backend. Its differentiators are support for any agent harness, sandbox, benchmark, and multiple backends through one API, plus token-faithful tracing and workflow-level training; this contrasts with more backend- or agent-shape-specific RL stacks.
Leadership
Executive Team
Sijun Tan
Project Lead
Fourth-year UC Berkeley CS PhD student advised by Raluca Ada Popa and affiliated with the Sky Computing Lab; previously earned CS and mathematics degrees at the University of Virginia, interned at Facebook AI Research, and worked as a senior algorithm engineer at Ant Group.
Tianhao Wu
Project Lead
Fifth-year UC Berkeley EECS PhD student advised by Jiantao Jiao and Kannan Ramchandran; studied mathematics at Peking University, previously researched instruction following, reasoning, and self-play RL, and is a core contributor to rLLM and the Hive project.
Board of Directors
Founding Story
rLLM began as an open-source initiative spawning from Berkeley's Sky Computing Lab to make reinforcement-learning techniques for LLMs and agents more accessible. The initial vision was to let researchers and developers build custom agents and environments, train them with RL, and use the same agent code for evaluation and training, avoiding the traditional inference-to-training rewrite.
Business Model
Revenue Model
The core rLLM framework is distributed as open-source software under the Apache License 2.0. The project also offers a hosted rLLM UI service alongside a self-hosted deployment option; public materials do not specify a paid pricing model.
Target Markets
- Academic AI/ML research labs
- Open-source model and agent researchers
- AI infrastructure and agent engineering teams
- Organizations training coding, research, finance, and other tool-using agents
- Developers using LangGraph, SmolAgents, Strands, OpenAI Agents SDK, Google ADK, or custom agent harnesses
- Post-training general-purpose and specialized language agents with reinforcement learning
- Long-horizon software-engineering, coding, and repository-migration agents
- Financial analysis and tool-using agents
- Deep-research and web-navigation agents
- Multi-agent solver-judge, self-verification, and agent-evolution workflows
- Evaluation and observability of agent trajectories in research and production-like deployments
- Alibaba NLP's Tongyi DeepResearch community project
- Snorkel AI collaboration on the FinQA financial-analysis agent and benchmark
- AWS AgentCore Science collaboration on Java repository migration agents
- Community projects including PettingLLMs, SETA, Terminal-Bench-RL, LLM-in-Sandbox, and V1 Parallel Self-Verification