Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning

What happened

Researchers from Renmin University, Ant Group, Tsinghua University and Zhejiang University published Ring-Zero (arXiv 2607.12395) — the first successful scaling of zero RL to 1 trillion parameters. With no human annotations, five emergent capabilities appeared spontaneously.

Context and impact

Zero RL (RL applied directly to a raw base model without a supervised fine-tuning phase) was considered infeasible at such large scales. Ring-Zero shows that at sufficient scale, capabilities emerge without human labeling — important for understanding future AGI-level system training.

Details

• Zero RL scaled to 1 trillion parameters — first time in history
• 5 emergent capabilities: self-verification, parallel reasoning, structured formatting, anthropomorphic narrative, context anxiety
• Context anxiety: model actively monitors and limits its own context window consumption in real time
• Skips the entire SFT phase — no human step-by-step annotations
• Confirms: emergence does not depend on human labeling at sufficient scale