All the AlphaZero offshoots, living under one PyTorch roof

What it does LightZero is a PyTorch toolkit that implements the major Monte Carlo Tree Search + Deep RL algorithms—AlphaZero, MuZero, EfficientZero, Gumbel MuZero, and several descendants—under one lightweight framework. It targets both board games (tic-tac-toe, gomoku, 2048) and continuous-control tasks (CartPole, LunarLander, Atari) so researchers can benchmark the entire algorithm family on equal footing. The project also ships with detailed algorithm diagrams, function-call graphs, and network-structure docs aimed at making the code understandable rather than merely runnable.

The interesting bit The authors treat standardization as a feature, not a side effect: every algorithm shares the same Model, Policy, and MCTS abstractions, and the hottest parts of the tree search are accelerated with a mixed Python/C++ backend. It is a NeurIPS 2023 Spotlight benchmark, and the team continues to publish derivative work (UniZero, ReZero, ScaleZero) that lands at ICLR and TMLR.

Key highlights

• Implements eight algorithms: AlphaZero, MuZero, Sampled MuZero, Stochastic MuZero, EfficientZero, Gumbel MuZero, ReZero, and UniZero.
• Uses heterogeneous computing (ptree in Python, ctree in C++) to speed up the most time-consuming MCTS operations.
• Provides algorithm framework diagrams, function call graphs, and network structure diagrams for every integrated method.
• NeurIPS 2023 Spotlight (Datasets and Benchmarks Track), with follow-up papers accepted at TMLR 2025, CoRL 2025, and ICLR 2026.
• Many algorithm-environment pairings remain locked; Chess and Go are not yet supported.

Caveats

• A large portion of the environment-algorithm matrix is marked with a 🔒, including full support for Chess and Go; many advanced algorithms only run on a subset of tasks.
• The README truncates the MuJoCo and DeepMind Control coverage table, so exact support there is unclear.

Verdict Researchers who need to compare MCTS-based RL methods head-to-head should keep this on their shortlist. If you just want a single algorithm running on a single game, the unified-abstraction overhead might feel like over-engineering.