Ring-Zero Scales Zero-RL to 1 Trillion Parameters — and Reports 'Context Anxiety' as an Emergent Behavior

arXiv / HuggingFace Daily Papers·medium signal

A 16-author team trained Ring-2.5-1T-Zero with pure zero-RL at trillion-parameter scale, finding that 'scaling to 1T parameters significantly enhances sample efficiency and performance ceilings,' with training splitting into a discovery phase followed by a sharpening phase. The model spontaneously develops anthropomorphism, structured formatting, self-verification, parallel reasoning, and 'context anxiety.' Competitive across seven math benchmarks, with the team introducing a chain-of-thought evaluation framework scoring comprehensibility, reproducibility, and efficiency. Getting there required clipped importance sampling, training-inference ratio correction, and mixed-precision control.