Researchers have demonstrated the effectiveness of scaling Zero RL (reinforcement learning with verifiable rewards without human-annotated data) up to one trillion parameters, revealing significant gains in model reasoning and sample efficiency. The team, including Xinyu Tang and colleagues, addressed common scaling challenges such as poor readability and token redundancy by implementing a specialized training pipeline. This pipeline utilizes key optimizations including clipped importance sampling, training-inference ratio correction, and mixed-precision control to maintain stability at large scale.

Experiments with the Ring-2.5-1T-Zero model identified three primary scaling effects. First, scaling to one trillion parameters enhances both performance ceilings and sample efficiency. Second, the training process unfolds in two distinct phases: an initial discovery phase followed by a sharpening phase. Third, the model spontaneously acquires advanced cognitive behaviors, such as self-verification, parallel reasoning, and structured formatting, which render manual heuristics unnecessary.

The study evaluated the model across seven mathematical benchmarks, achieving competitive performance. Beyond final-answer accuracy, the researchers introduced a structured framework to assess Chain-of-Thought (reasoning steps explaining the path to a solution) quality across three dimensions: comprehensibility, reproducibility, and efficiency. Results indicate the one-trillion-parameter scale provides clear advantages in generating concise, structured reasoning traces, offering new insights into how large-scale models develop emergent capabilities.

Researchers have demonstrated the effectiveness of scaling Zero RL (reinforcement learning with verifiable rewards without human-annotated data) up to one trillion parameters, revealing significant gains in model reasoning and sample efficiency. The team, including Xinyu Tang and colleagues, addressed common scaling challenges such as poor readability and token redundancy by implementing a specialized training pipeline. This pipeline utilizes key optimizations including clipped importance sampling, training-inference ratio correction, and mixed-precision control to maintain stability at large scale.

Experiments with the Ring-2.5-1T-Zero model identified three primary scaling effects. First, scaling to one trillion parameters enhances both performance ceilings and sample efficiency. Second, the training process unfolds in two distinct phases: an initial discovery phase followed by a sharpening phase. Third, the model spontaneously acquires advanced cognitive behaviors, such as self-verification, parallel reasoning, and structured formatting, which render manual heuristics unnecessary.

The study evaluated the model across seven mathematical benchmarks, achieving competitive performance. Beyond final-answer accuracy, the researchers introduced a structured framework to assess Chain-of-Thought (reasoning steps explaining the path to a solution) quality across three dimensions: comprehensibility, reproducibility, and efficiency. Results indicate the one-trillion-parameter scale provides clear advantages in generating concise, structured reasoning traces, offering new insights into how large-scale models develop emergent capabilities.