A reinforcement learning (RL) infrastructure designed to bridge foundation model training with modern agent-based applications.
We are part of the PyTorch Foundation Landscape Project.
A microservice-based redesign of AReaL that decouples training, inference, agent execution and weight-update into independent services — enabling long-horizon, self-evolving agents at scale.
AReaL-SEA, a self-evolving data-synthesis engine paired with RL training. A 235B MoE surpasses GPT-5 and matches Gemini 3.0 Pro on τ²-bench.
A DFS-based dynamic tree attention that exploits shared rollout prefixes during RL post-training — up to 8.31× throughput over dense training on τ²-bench, with a load-balanced distributed batching scheme across GPUs.
The original AReaL system paper — a fully asynchronous training paradigm that achieves 2.77× speedup while matching or exceeding synchronous baselines on reasoning benchmarks.
