Ring-Zero: Scaling Zero RL to a Trillion Parameters

With Renmin University's Gaoling School (plus Tsinghua and Zhejiang): zero RL — RLVR directly from a pretrained base model, no SFT or human-annotated CoT — scaled to Ling-2.5-1T-Base (1T total / 63B active MoE) and Ling-2.5-flash-Base (104B / 7.4B). A four-phase pipeline (clipped importance sampling, training-inference ratio correction, mixed-precision control, self-distillation) tames readability and token redundancy at scale. Findings: 1T scale markedly improves sample efficiency and performance ceilings; training follows a discovery phase then a sharpening phase; and advanced behaviors emerge spontaneously — self-verification, parallel reasoning, structured formatting, anthropomorphism, and "context anxiety". Evaluated on seven competition math benchmarks against frontier models. Paper-only release.