Ring-Zero: Scaling Zero RL to 1 Trillion Parameters with Emergent Reasoning Behaviors
Ant Group
Ring-Zero trains a 1-trillion-parameter language model using reinforcement learning with verifiable rewards and zero human-annotated chain-of-thought data. The pipeline introduces clipped importance sampling, training-inference ratio correction, and mixed-precision control to stabilize optimization at this scale. Ring-2.5-1T-Zero achieves 84.2% on AIME 2026 without SFT warm-up. The model spontaneously develops structured formatting, self-verification, parallel reasoning, and what the authors call 'context anxiety' — emergent behaviors not hand-coded into the training process.
Why it matters
First publicly documented application of zero-RL training at the trillion-parameter scale, revealing that scale alone induces qualitatively new reasoning behaviors and removes the need for hand-crafted chain-of-thought heuristics. Ranked 2nd on HuggingFace Daily Papers (61 upvotes).
Importance: 4/5
First trillion-parameter zero-RL training; emergent cognitive behaviors at scale; 61 HF upvotes; 84.2% on AIME 2026 without SFT