Flax is a neural network library for JAX that is designed for flexibility.

Mastering Diverse Domains through World Models

Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos

Benchmarking the Spectrum of Agent Capabilities

Evaluating long-term memory of reinforcement learning algorithms

Decoupled Q-Chunking

(Crafter + NetHack) in JAX. ICML 2024 Spotlight.

A framework for creating rich, 3D, Minecraft-like single and multi-agent environments for AI research. (Accepted at ICML 2025).

Code for "Transitive RL: Value Learning via Divide and Conquer"

The first behavioral foundation model to control a virtual physics-based humanoid agent for a wide range of whole-body tasks.

Code for "Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining"

Training teachers with reinforcement learning able to make LLMs learn how to reason for test time scaling.

JAX implementation of VQVAE/VQGAN autoencoders (+FSQ)

[EMNLP 2025 Main] AlphaOne: Reasoning Models Thinking Slow and Fast at Test Time

Code for A Multi-Region Brain Model to Elucidate the Role of Hippocampus in Spatially Embedded Decision-Making (ICML 2025)

[NeurIPS 2025 Spotlight] Reasoning Environments for Reinforcement Learning with Verifiable Rewards

The official implementation of "Horizon Reduction Makes RL Scalable"

Fine-tuning & Reinforcement Learning for LLMs. 🦥 Train OpenAI gpt-oss, DeepSeek, Qwen, Llama, Gemma, TTS 2x faster with 70% less VRAM.

Extreme Q-Learning: Max Entropy RL without Entropy

PyTorch code for Vision Transformers training with the Self-Supervised learning method DINO

Code for "Unsupervised Zero-Shot RL via Functional Reward Representations"

Foundation Policies with Hilbert Representations (ICML 2024)