XiaomiMiMo/verl
Agentic RL training code for MiMo. The detailed training recipe can be found in Section 7 of our report MiMo-V2.6: Scaling Reinforcement Learning Towards
Self-Improvement.
This fork adds reproduction code for five RL environments upon verl (0.9.0.dev).
| Domain |
Task Family |
Verifier |
Launch script |
Settings |
| Code |
Software engineering |
Executable tests |
scripts/code/train.sh |
scripts/code/env.example |
| Cyber |
Vulnerability reproduction |
Rule checks |
scripts/arvo/arvo.sh |
scripts/arvo/arvo.env.example |
| General |
Knowledge work |
Rubric-based judging |
scripts/general/general.sh |
scripts/general/general.env.example |
| Visual |
Web development |
Visual grading |
scripts/design/webdev.sh |
scripts/design/webdev.env.example |
| Music |
Symbolic music composition |
Rule checks |
scripts/design/music.sh |
scripts/design/music.env.example |
Each launch script reads the variables listed in the *.env.example next to it; set those before launching.
The original verl README follows.
👋 Hi, everyone!
verl is a RL training library initiated by ByteDance Seed team and maintained by the verl community.
verl: Volcano Engine Reinforcement Learning for LLMs
verl is a flexible, efficient and production-ready RL training library for large language models (LLMs).
verl is the open-source version of HybridFlow: A Flexible and Efficient RLHF Framework paper.
verl is flexible and easy to use with:
Easy extension of diverse RL algorithms: The hybrid-controller programming model enables flexible representation and efficient execution of complex post-training dataflows. Build RL dataflows such as GRPO, PPO in a few lines of code.
Seamless integration of existing LLM infra with modular APIs: Decouples computation and data dependencies, enabling seamless integration with existing LLM frameworks, such as FSDP, Megatron-LM, vLLM, SGLang, etc
Flexible device mapping: Supports various placement of models onto different sets of GPUs for efficient resource utilization and scalability across different cluster sizes.
Ready integration with popular HuggingFace models
verl is fast with:
State-of-the-art throughput: SOTA LLM training and inference engine integrations and SOTA RL throughput.
Efficient actor model resharding with 3D-HybridEngine: Eliminates memory redundancy and significantly reduces communication overhead during transitions between training and generation phases.
News
- [2026/07] RL-Insight is released: online observability for reinforcement learning training. RL-Insight connects training-side metrics, RL state traces, and service dashboards across distributed rollout and optimization workloads.
- [2026/06] verl-SpeCo is pre-released: a co-training framework for speculative decoding across RL training and inference, keeping draft models aligned during training and reusable for accelerated serving, built on top of verl.
- [2026/05] uni-agent is released: a unified agent framework to build, run, and train LLM agents at scale, built on top of verl.
- [2026/05] VeRL-Omni is pre-released: a unified RL stack for diffusion and omni-modal model post-training built on top of verl. Read the blog post for details.
- [2026/05] verl's zero-mismatch HuggingFace rollout vexact is released: with batch-invariant kernels, shared model definition with FSDP, and out-of-box examples compatible with VeOmni.
- [2026/04] verl's Megatron backend LoRA and router replay support is showcased at PyTorch Conference Europe 2026.
- [2026/03] verl is presented at NVIDIA GTC26: session#1, session#2
- [2026/01] verl has been migrated to the verl-project
- [2026/01] verl first meetup was successfully held in Shanghai on 01/10, hosted by Volcengine and NVIDIA, the slides has been uploaded to verl-data.
- [2026/01] The
recipe directory has been migrated to a dedicated repository: verl-recipe and added as a submodule. See https://github.com/verl-project/verl/pull/4795. It can be used as it was after git submodule update --init --recursive recipe. Note that transfer_queue, fully_async_policy, one_step_off_policy and vla are kept under verl/experimental since they are planned to be merged into the main library. Use them through verl.experimental.{module}.
- [2025/12] Mind Lab successfully used verl and Megatron-bridge to train GRPO Lora for Trillion-parameter model on 64 H800 - See their techblog.
- [2025/10] verl is presented in the PyTorch Conference 2025.
- [2025/08] verl is presented in the PyTorch Expert Exchange Webinar. Slides available.
- [2025/07] The ReTool recipe is fully open sourced. Blog
- [2025/07] The first verl meetup will be held at ICML Vancouver on July 16th! Please join us if you are at ICML! (onsite only)
- [2025/06] verl with Megatron backend enables large MoE models such as DeepSeek-671B and Qwen3-235B.
- [2025/03] DAPO is the open-sourced SOTA RL algorithm that achieves 50 points on AIME 2024 based on the Qwen2.5-32B pre-trained model, surpassing the previous SOTA achieved by DeepSeek's GRPO (DeepSeek-R1-Zero-Qwen-32B). DAPO's training is fully powered by verl and the reproduction code is available in
recipe/dapo now.
more...
- [2025/04] [Seed-Thinking-v1.5](https://github.com/ByteDance-Seed/Seed-Thinking-v1.5/blob/main/seed-thinking-v1.5.pdf) tech report is released! Trained with verl, Seed-Thinking-v1.5 achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains.
- [2025/07] verl keynote at [AWS AI Hours Singapore](https://pages.awscloud.com/aws-ai-hours-sg.html#agenda) on 7/8, verl & verl-agent project updates at [Agent for SWE meetup](https://lu.ma/e498qhsi) by LF AI & Data Singapore on 7/11.
- [2025/06] verl team will provide latest project updates at [PyTorch Day China](https://www.lfasiallc.com/pytorch-day-china/) on June 7th. Meet our dev team in Beijing!
- [2025/04] [VAPO](https://arxiv.org/pdf/2504.05118) (value-based augmented PPO) paper covers our latest RL method for reasoning models. Trained from Qwen-32B-base model, VAPO achieves 60.4 on AIME 2024, outperforming DAPO-32B.
- [2025/05] [PF-PPO](https://arxiv.org/abs/2409.06957), accepted to ICML 2025, is now supported in verl! PF-PPO enhances policy learning efficiency and robustness by filtering potentially noisy reward signals and reusing high-quality experiences via a replay buffer.
- [2025/04] We will give a tutorial about latest post-training techniques and programming guide for verl at [ICLR 2025 Expo](https://iclr.cc/virtual/2025/calendar?filter_events=Expo+Talk+Panel&filter_rooms=), [SCI-FM workshop](https://open-foundation-model.github.io/) and [LMSys afterparty](https://lu.ma/d23nyynm). Talk materials available [here](https://github.com/eric-haibin-lin/verl-community/tree/main/iclr25).
- [2025/03] verl v0.3.0.post1 is released! See [release note](https://github.com/verl-project/verl/releases/) for details. It achieves [~1.4x speedup](https://tongyx361.github.io/blogs/posts/verl-intro/#/verl-flexible-and-efficient-rl-for-llms) compared to prev versions.
- [2025/05] verl will be presented at [A2M Shanghai](https://a2m.msup.com.cn/home/?aid=4488&city=shanghai) on 5/16 - 5/17.
- [2025/05] verl will be presented at [GOSIM x PyTorch Day 2025](https://paris2025.gosim.org/). See you in Paris!
- [2025/03] We introduced the programming model of verl at the [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg) and [verl intro and updates](https://github.com/eric-haibin-lin/verl-community/blob/main/slides/verl-lmsys-meetup.pdf) at the [SGLang-LMSYS Org Meetup](https://lu.ma/ntjrr7ig) in Sunnyvale mid-March.
- [2025/03] We will present verl(HybridFlow) at EuroSys 2025. See you in Rotterdam!
- [2025/02] verl v0.2.0.post2 is released!
- [2025/02] We presented verl in the Bytedance/NVIDIA/Anyscale Ray Meetup. See you in San Jose!
- [2025/01] [Doubao-1.5-pro](https://team.doubao.com/zh/special/doubao_1_5_pro) is released with SOTA-level performance on LLM & VLM. The RL scaling preview model is trained using verl, reaching OpenAI O1-level performance on math benchmarks (70.0 pass@1 on AIME).
- [2024/12] verl is presented at Ray Forward 2024. Slides available here
- [2024/12] The team presented Post-training LLMs: From Algorithms to Infrastructure at NeurIPS 2024. Slides and video available.
- [2024/10] verl is presented at Ray Summit. Youtube video available.
- [2024/08] HybridFlow (verl) is accepted to EuroSys 2025.
Key Features
- FSDP, FSDP2 and Megatron-LM for training.
- vLLM, SGLang and HF Transformers for rollout generation.
- Compatible with Hugging Face Transformers and Modelscope Hub: Qwen3.5, Qwen3, Qwen-2.5, Llama3.1, Gemma2, DeepSeek-LLM, etc
- Supervised fine-tuning.
- Reinforcement learning with PPO, GRPO, GSPO, ReMax, REINFORCE++, RLOO, PRIME, DAPO, DrGRPO, KL_Cov & Clip_Cov etc.
- LLM alignment recipes such as Self-play preference optimization (SPPO)
- Flash attention 2, sequence packing, sequence parallelism via DeepSpeed Ulysses, LoRA, Liger-kernel (
USE_LIGER=1).
- Scales up to 671B models and hundreds of GPUs with expert parallelism
- Multi-gpu LoRA RL support to save memory.
- Experiment tracking with wandb, swanlab, mlflow and tensorboard.
- Hardware Support: Supports NVIDIA, AMD, Ascend
Getting Started
Documentation
Quickstart:
Running a PPO example step-by-step:
Reproducible algorithm baselines:
Algorithm recipes (recipe/):
- Optional workflows and baselines live under
recipe/. Each recipe subdirectory includes a small REQUIRED_VERL.txt file describing the intended verl install: pinned recipes use a tag or fixed git SHA; rolling recipes record an explicit VERL_COMMIT (and related submodule / recipe-folder SHAs) so you can pip install verl@git+…@<sha> without guessing. See recipe/README.md for the full index and links.
For code explanation and advance usage (extension):
Blogs from the community
The performance is essential for on-policy RL algorithm. We have written a detailed performance tuning guide to help you optimize performance.
Upgrade to vLLM >= v0.8.2
verl now supports vLLM>=0.8.2 when using FSDP as the training backend. Please refer to this document for the installation guide and more information. Please avoid vllm 0.7.x, which contains bugs that may lead to OOMs and unexpected errors.
Use Latest SGLang
SGLang is fully supported with verl, and SGLang RL Group is working extensively on building unique features, including multi-turn agentic RL, VLM RLHF, server-based RL, and partial rollout. Please refer to this document for the installation guide and more information.
Upgrade to FSDP2
verl is fully embracing FSDP2! FSDP2 is recommended by torch distributed team, providing better throughput and memory usage, and is composible with other features (e.g. torch.compile). To enable FSDP2, simply use verl main and set the following options:
actor_rollout_ref.ref.strategy=fsdp2
actor_rollout_ref.actor.strategy=fsdp2
critic.strategy=fsdp2
Furthermore, FSDP2 cpu offloading is compatible with gradient accumulation. You can turn it on to save memory with actor_rollout_ref.actor.fsdp_config.offload_policy=True. For more details, see https://github.com/verl-project/verl/pull/1026
AMD Support (ROCm Kernel)
verl runs on AMD ROCm GPUs (MI300X / MI325X / MI355X) with FSDP, FSDP2, and Megatron trainer backends, and vLLM as the validated inference engine (SGLang support is in progress). See the AMD ROCm quick-start guide for container bring-up, environment verification, and training examples.
Citation and acknowledgement
If you find the project helpful, please cite:
@article{sheng2024hybridflow,
title = {HybridFlow: A Flexible and Efficient RLHF Framework},
author = {Guangming Sheng and Chi Zhang and Zilingfeng Ye and Xibin Wu and Wang Zhang and Ru Zhang and Yanghua Peng and Haibin Lin and Chuan Wu},
year = {2024},
journal = {arXiv preprint arXiv: 2409.19256}
}
verl is inspired by the design of Nemo-Aligner, Deepspeed-chat and OpenRLHF. The project is adopted and contributed by Bytedance, Anyscale, LMSys.org, Alibaba Qwen team, Shanghai AI Lab, Tsinghua University, UC Berkeley, UCLA, UIUC, University of Hong Kong, ke.com, All Hands AI, ModelBest, JD AI Lab, Microsoft Research, StepFun, Amazon, LinkedIn, Meituan, Camel-AI, OpenManus, Xiaomi, NVIDIA research, Baichuan, RedNote, SwissAI, Moonshot AI (Kimi), Baidu, Snowflake, Skywork.ai, JetBrains, IceSword Lab, and many more.
Awesome Projects Built with verl
Welcome to register your awesome project build with verl for other developers' reference!
- TinyZero: a reproduction of DeepSeek R1 Zero recipe for reasoning tasks
- SkyThought: RL training for Sky-T1-7B by NovaSky AI team.
- simpleRL-reason: SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild
- Easy-R1: Multi-modal RL training framework
- RandOpt: Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights
- OpenManus-RL: LLM Agents RL tuning framework for multiple agent environments.
- rllm: async RL training with verl-pipeline
- RAGEN: a general-purpose reasoning agent training framework
- Search-R1: RL with reasoning and searching (tool-call) interleaved LLMs
- ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning
- Skywork-OR1: Skywork open reaonser series
- ToRL: Scaling tool-integrated RL
- Absolute Zero Reasoner: A no human curated data self-play framework for reasoning
- verl-agent: A scalable training framework for long-horizon LLM/VLM agents, along with a new algorithm GiGPO
- RL-Factory: An easy and efficient RL post-training framework for Agentic Learning
- ReTool: ReTool: reinforcement learning for strategic tool use in LLMs. Code release is in progress...
- verl-tool: An unified and easy-to-extend tool-agent training framework based on verl
- PRIME: Process reinforcement through implicit rewards
- MemAgent: MemAgent: Reshaping Long-Context LLM with Multi-Conv RL based Memory Agent
- POLARIS: A Post-training recipe for scaling RL on Advanced Reasoning models
- GUI-R1: GUI-R1: A Generalist R1-style Vision-Language Action Model For GUI Agents
- DeepRetrieval: RL Training of Search Agent with Search/Retrieval Outcome
- Code-R1: Reproducing R1 for Code with Reliable Rewards
- DeepResearcher: Scaling deep research via reinforcement learning in real-world environments
- VAGEN: Training VLM agents with multi-turn reinforcement learning
- RM-R1: RL training of reasoning reward models
- Dr. MAS: Stable end-to-end RL post-training for multi-agent LLM systems
- LUFFY: Learning to Reason under Off-Policy Guidance
- DeepMath: DeepMath-103K data and series models for math reasoning
- PACS: Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR
- Entropy Mechanism of RL: The Entropy Mechanism of Reinforcement Learning for Large Language Model Reasoning
- LLaSA-TTS-GRPO: TTS fine-tuning with GRPO optimization based on LLASA models
- PF-PPO: Policy Filtration for PPO based on the reliability of reward signals for more efficient and robust RLHF.
- RACRO: Build multi-modal reasoning models via decoupling it into query-conditioned captioning and text-only reasoning
- Agent Lightning: A flexible and extensible framework that enables seamless agent optimization for any existing agent framework.
- VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use.
- Kimina-Prover-RL: Training pipeline for formal theorem proving, based on a paradigm inspired by DeepSeek-R1.
- RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization.
- rStar2-Agent: Using reinforcement learning with multi-step tool-calling for math tasks, rStar2-Agent-14B reaches frontier-level math reasoning in just 510 RL training steps
- Vision-SR1: Self-Rewarding Vision-Language Model via Reasoning Decomposition
- SimpleVLA-RL: SimpleVLA-RL: A Simple yet Effective Vision-Language Action Model for Reinforcement Learning
- Table-R1: Table-R1: Inference-Time Scaling for Table Reasoning
- Revisual-R1: Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning
- ARES: ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping
- Meta-Bandit-LLM: Meta-Bandit-LLM: Long-horizon multiturn interactive training for meta-bandit agents
- PokeeResearch: PokeeResearch: State-of-the-art 7B DeepResearch Agent that leverages web search and content reading capabilities to answer complex questions using the most up-to-date information available online.
- Search Self-play: Pushing the Frontier of Agent Capability without Supervision
- OneThinker: All-in-one Reasoning Model for Image and Video
- OpenTinker: Democratizing Agentic Reinforcement Learning as a Service
- FlowRL: Matching reward distributions via flow balance for diverse exploration and generalizable reasoning
- Logic-RL: a reproduction of DeepSeek R1 Zero on 2K Tiny Logic Puzzle Dataset.
- Seed-Coder: RL training of Seed-Coder boosts performance on competitive programming
- all-hands/openhands-lm-32b-v0.1: A strong, open coding agent model, trained with multi-turn fine-tuning
- s3 Efficient Yet Effective Search Agent Training via RL
- Rec-R1: Bridging Generative Large Language Models and Recommendation Systems via Reinforcement Learning
- Explore RL Data Scaling: Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback
- FIRE: Flaming-hot initiation with regular execution sampling for large language models
- DQO: Enhancing multi-Step reasoning abilities of language models through direct Q-function optimization
- ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models
- cognition-engineering: Test time scaling drives cognition engineering.
- Trust Region Preference Approximation: A simple and stable reinforcement learning algorithm for LLM reasoning.
- AdaRFT: Efficient Reinforcement Finetuning via Adaptive Curriculum Learning
- critic-rl: LLM critics for code generation
- self-rewarding-reasoning-LLM: self-rewarding and correction with generative reward models
- DeepEnlighten: Reproduce R1 with social reasoning tasks and analyze key findings
- MetaSpatial: Reinforcing 3D Spatial Reasoning in VLMs for the Metaverse
- PURE: Credit assignment is the key to successful reinforcement fine-tuning using process reward model
- cognitive-behaviors: Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs
- deepscaler: iterative context scaling with GRPO
- DAPO: the fully open source SOTA RL algorithm that beats DeepSeek-R1-zero-32B
- NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation
- SPEAR: Self-imitation with Progressive Exploration for Agentic Reinforcement Learning (ICLR 2026)
- RuleReasoner: RuleReasoner: Reinforced Rule-based Reasoning via Domain-aware Dynamic Sampling (ICLR 2026)
- MetaphorStar: Image Metaphor Understanding and Reasoning with End-to-End Visual Reinforcement Learning
- DART-GUI: a decoupled agentic RL framework for Computer Use Agents, achieving ~2× training speedup and ~5× environment utilization!
- Rethinking OPD: Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
Contribution Guide
See contributions guide
Founded in 2023, ByteDance Seed Team is dedicated to crafting the industry's most advanced AI foundation models. The team aspires to become a world-class research team and make significant contributions to the advancement of science and society. You can get to know Bytedance Seed better through the following channels👇
We are HIRING! Send us an email if you are interested in internship/FTE opportunities in RL for agents.
verl is an open-source RL post-training framework for large language models, started by ByteDance Seed as the open-source version of the HybridFlow paper. This XiaomiMiMo fork adds agentic RL training code for MiMo, with reproduction code for five RL environments on top of verl 0.9.0.dev.