Scaling Zero RL to One Trillion Parameters Unlocks Emergent Reasoning

Researchers have demonstrated that scaling reinforcement learning without human-annotated data—known as "zero RL"—to a trillion parameters can elicit advanced reasoning capabilities. According to a paper published via arXiv, the resulting model, Ring-2.5-1T-Zero, shows that scaling significantly boosts performance ceilings and sample efficiency.

As reported by arXiv, the authors—including Xinyu Tang, Gangqiang Cao, Yurou Liu, and 13 others—found that the training process moves through two distinct stages: an initial discovery phase and a subsequent sharpening phase. To reach this scale, the team implemented a training pipeline featuring mixed-precision control, training-inference ratio correction, and clipped importance sampling to overcome issues with token redundancy and poor readability.

The findings suggest that at the 1-trillion parameter scale, models spontaneously develop cognitive behaviors that make hand-crafted heuristics unnecessary. These emergent behaviors include parallel reasoning, self-verification, structured formatting, anthropomorphism, and "context anxiety."

Evaluated across seven mathematical benchmarks, Ring-2.5-1T-Zero achieved competitive results. The researchers also introduced a new evaluation framework to measure the quality of chain-of-thought reasoning based on efficiency, reproducibility, and comprehensibility, noting that the model produces reasoning traces that are more concise and structured.