OpenDILab @OpenDILab

Excited to share ScaleZero (ICLR 2026)! 🚀 We propose "One Model for All Tasks" using MoE & Dynamic Parameter Scaling to mitigate gradient conflicts and boost sample efficiency in heterogeneous multi-task RL. 🔗 arxiv.org/abs/2509.07945 💻 github.com/opendilab/Ligh…

keyboard_arrow_left Previous keyboard_arrow_right Next

Key highlights: ✅ Only 24 tokens per second for 24kHz audio ✅ Single-quantizer inference ✅ State-of-the-art speech reconstruction ✅ Optimized for spoken language modeling

🚀 Excited to share our new work: HH-Codec, a neural speech codec that achieves ultra-low bitrate (0.3 kbps) while preserving high-fidelity reconstruction! 🎧 paper：arxiv.org/pdf/2507.18897 repo：github.com/opendilab/HH-C… #NeuralCodec #AudioAI #SpeechProcessing #AI

We just launched awesome-RLVR—a curated hub for Reinforcement Learning with Verifiable Rewards (RLVR), featuring tutorials, codebases, and papers! Stay tuned for updates with cutting-edge research & open-source contributions. Star, follow & share! github.com/opendilab/awes…

👏We are delighted to announce the release of LightZero v0.2.0! This version brings a series of enhancements and fixes some known issues. You can find more details on github.com/opendilab/Ligh… Thank you for your support of LightZero!

🚀 Exciting news! Our research paper: Empowering LLMs in Decision Games through Algorithmic Data Synthesis has been accepted to the ICLR 2025 Workshop SynthData! It highlights how decision-game data can serve as reasoning cases for LLMs! Arxiv: arxiv.org/abs/2503.13980

👏We are delighted to announce the release of DI-engine v0.5.3! This version brings a series of enhancements and fixes some known issues. You can find more details on this link (github.com/opendilab/DI-e…). Thank you for your support of DI-engine!

🎉 We are thrilled to announce that our GitHub repository github.com/opendilab/DI-e… has reached 3k stars! Thank you all for your support! It is a python-first and asynchronous-native task Reinforcement Learning framework, modularly integrating 60+ RL envs and 80+ RL algorithms.

@such_a_sophist Hi Sophie, I'm glad to hear from you! Here's the invitation link to our Discord server: discord.com/invite/dkZS2JF… Looking forward to seeing you there!

We create our discord channel today, welcome to more discussions and feedbacks. Here is the invitation link: discord.gg/TC2pAtnw

🎉 We are thrilled to announce that our GitHub repository github.com/opendilab/Ligh… has reached 1k stars! Thank you all for your support! We will continue to promote the standardization of the MCTS+RL algorithm family to accelerate related research and applications. 🚀

AMEX Android Multi-annotation Expo Dataset for Mobile GUI Agents AI agents have drawn increasing attention mostly on their ability to perceive environments, understand tasks, and autonomously achieve goals. To advance research on AI agents in mobile scenarios, we introduce the Android Multi-annotation EXpo (AMEX), a comprehensive, large-scale dataset designed for generalist mobile GUI-control agents. Their capabilities of completing complex tasks by directly interacting with the graphical user interface (GUI) on mobile devices are trained and evaluated with the proposed dataset. AMEX comprises over 104K high-resolution screenshots from 110 popular mobile applications, which are annotated at multiple levels. Unlike existing mobile device-control datasets, e.g., MoTIF, AitW, etc., AMEX includes three levels of annotations: GUI interactive element grounding, GUI screen and element functionality descriptions, and complex natural language instructions, each averaging 13 steps with stepwise GUI-action chains. We develop this dataset from a more instructive and detailed perspective, complementing the general settings of existing datasets. Additionally, we develop a baseline model SPHINX Agent and compare its performance across state-of-the-art agents trained on other datasets. To facilitate further research, we open-source our dataset, models, and relevant evaluation tools.

（continued） - [DI-engine](github.com/opendilab/DI-e…): Our comprehensive RL library - [PPOxFamily](github.com/opendilab/PPOx…): Our beginner's guide to RL - [LightZero](github.com/opendilab/Ligh…): Our MCTS + RL framework Together, we continue to strive towards AGI! 🚀

🌟 Exciting News! 🌟 Our GitHub organization [OpenDILab](github.com/opendilab) has reached an incredible milestone of 20k stars! ⭐️ A huge thank you to everyone for your support! 🙏 Special shoutouts to our key repositories: （to be continued）

We are excited to announce our latest research paper, 'UniZero: Generalized and Efficient Planning with Scalable Latent World Models,' now available at arxiv.org/abs/2406.10667, i.e. a scalable model-policy joint learning approach to enhance planning efficiency and generality.

keyboard_arrow_left Previous keyboard_arrow_right Next

👏 We are excited to announce the release of LightZero v0.1.0! This version introduces two innovative algorithms: UniZero and ReZero, and addresses several known issues. For more details, please visit github.com/opendilab/Ligh…. Thank you for your continued support of LightZero!

We are delighted to announce the release of Awesome UI Agent! A collection of research papers for UI Agent, which includes models, tools, and datasets. Github:github.com/opendilab/awes… Try it now!

👏👏👏We are delighted to announce the release of MasterMind dataset! The expert dataset for the Doudizhu and Go tasks proposed in MasterMind. Here are corresponding links: HuggingFace: huggingface.co/datasets/OpenD… OpenDataLab: opendatalab.com/OpenDILab/Mast… Try it now!😍😍😍😍

keyboard_arrow_left Previous keyboard_arrow_right Next

We are excited to release our latest research paper titled "ReZero: Boosting MCTS-based Algorithms by Just-in-Time and Speedy Reanalyze" on arxiv.org/pdf/2404.16364. ReZero is designed to enhance the training speed of MCTS-based algorithms while maintaining high sample efficiency.

keyboard_arrow_left Previous keyboard_arrow_right Next

👏We are delighted to announce the release of LightZero v0.0.5! This version brings a series of enhancements and fixes some known issues. You can find more details on github.com/opendilab/Ligh…. Thank you for your support of LightZero!