Ring-Zero Paper Trains 1T-Parameter AI Reasoner Using RL Without Human Examples

A new arXiv paper titled 'Ring-Zero' presents a one-trillion-parameter mixture-of-experts model trained entirely through reinforcement learning from verifiable rewards, without relying on human-written reasoning traces. The model, developed using 320 H200 GPUs, achieved 84.2% pass@1 on the AIME 2026 benchmark at its first training stage, with scores climbing into the low 90s on several math benchmarks after later stages. Training followed a four-step pipeline including two RL phases — an early discovery phase to surface new reasoning behaviors and a later refinement phase to sharpen the policy. A notable feature is the model's tiered inference system, allowing it to allocate different token budgets of 4k, 16k, or 64k depending on problem complexity, rather than applying maximum compute to every query. The paper argues that larger models trained this way are more sample-efficient, and that emergent behaviors such as self-verification and structured formatting arose naturally from the training dynamics rather than explicit human instruction.