Papers.Data.Code
Paper

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

Xinyu Tang, Gangqiang Cao, Yurou Liu, Yuliang Zhan, Xiaochong Lan, Yifan Li, Yuchen Yan, Han Peng

A staged zero-RL pipeline for trillion-scale reasoning that replaces heuristic CoT shaping with stabilized RL, self-distillation, and adaptive-depth training.

arXiv ↗Jul 14, 2026

  • nlp
  • llm
  • reinforcement learning
  • reasoning
  • chain-of-thought
Multi-stage training pipeline diagram with infrastructure and emergent behaviors overview
Multi-stage training pipeline diagram with infrastructure and emergent behaviors overview

What it is

Scales zero-data RL with verifiable rewards to a 1T-parameter LLM using a minimal, stabilized pipeline for concise, high-quality reasoning. Instead of adding hand-crafted reasoning heuristics, it scales zero RL to 1T with a simple staged pipeline: clipped IS with train-infer ratio correction, self-distillation to compress CoT, then sample-level and tiered RL for adaptive depth.

Key results

  • 1T model shows higher sample efficiency than its 104B counterpart
  • Competitive performance on 7 challenging mathematical benchmarks
  • Better CoT quality with clearer, more concise traces under 3-dim eval

At trillion scale, simple stabilized zero-RL can elicit stronger and more structured reasoning than hand-crafted prompting heuristics.

One useful ML email a week.

More from Jul 13 – Jul 19

The whole week →

Picked Jul 17

Picked Jul 16

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.