$ cat wiki/papers/2026/2607.12395-ring-zero-rlvr-1t.md
Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning
TL;DR
First demonstration of RLVR (RL with Verifiable Rewards) scaling to 1 trillion parameters — achieves 84.2% on AIME 2026 and documents five emergent behaviors that don't appear at smaller scales, including a novel "context anxiety" failure mode.
Authors & Org
Xinyu Tang, Qianggang Cao, Yurou Liu, and 13 others — Ant Group (InclusionAI team)..
Method
Scales the "zero RL" paradigm (minimal inductive bias, no human labels) to 1T parameters using only verifiable reward signals from math/coding benchmarks. No supervised fine-tuning prior; the model learns purely from correctness feedback.
Key engineering: Ring-Zero refers to a training architecture (details in paper) enabling stable RLVR at 1T scale — prior work stalled below 100B.
Results
- AIME 2026: 84.2% — strong math reasoning result at this scale
- Five emergent behaviors documented:
- Self-verification: model checks its own intermediate answers before committing
- Context anxiety: hedges or stalls when the context is ambiguous — a novel failure mode not present below ~100B
- Hierarchical planning: spontaneous decomposition into sub-steps
- (Two additional behaviors noted in TechTimes coverage; full list in paper)
Significance
Closes a significant empirical gap: RLVR scaling laws now have a 1T data point. The "context anxiety" emergence is the most alarming finding — a qualitatively new behavior where larger models become less decisive under ambiguity rather than more, suggesting that naive RLVR scaling doesn't monotonically improve robustness. Relevant to alignment (unexpected emergent behaviors at scale).
Open Questions
- Is context anxiety a RLVR artifact or does it appear with SFT+RL pipelines too?
- How does Ring-Zero compare to test-time-compute approaches (e.g., best-of-N, verifier reranking) at the same parameter count?
- Can the five emergent behaviors be replicated across different model families?
Cite
arXiv:2607.12395 (~2026-07-12). HuggingFace Daily Papers ~July 14–16, 2026.
Related
- Reasoning Models — RLVR is the primary training paradigm for reasoning models
- Test-Time Compute (Inference-Time Compute Scaling) — complementary scaling axis
- Agentic Reinforcement Learning — agentic RL is the applied direction of RLVR
- SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning — SEED (same week; on-policy distillation for agents)
- Weak-to-Strong Generalization via Direct On-Policy Distillation — Direct-OPD (RLVR cost reduction at frontier scale)
- The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning — MIPI (RL training-inference probability mismatch)