InclusionAI's Ring-Zero Recharacterizes the Human-Data Ceiling as an Engineering One

The argument that capability training has hit a hard ceiling without human-curated reasoning data just lost a load-bearing wall. InclusionAI and Ant Group, in a paper posted July 14, 2026, report a model called Ring-2.5-1T-Zero trained using only reinforcement learning on automatically-checked answers. No human-written labels, no curated traces. The line that DeepSeek R1-Zero opened in January 2025, lifting AIME 2024 accuracy from 15.6% to above 70% on much smaller models, Ring-Zero extends to a scale that previously collapsed under naive training.

The paper's actual claim is not the parameter count. It is that the failure modes that killed zero RL above a few hundred billion parameters (poor readability, token redundancy, flat adaptive reasoning depth) are now addressable with three named engineering fixes: clipped importance sampling, training-inference ratio correction, and mixed-precision control. The five emergent behaviors the team names (self-verification, parallel reasoning, structured formatting, anthropomorphism, and "context anxiety") appear only after training passes through a measured discovery phase into a sharpening phase.

That last detail is the reusable one. The paper argues that engineering the verifier, the ratio, and the precision is what opened the path to 1T scale, not the parameter count itself. The reader should expect the next papers in this line to attack the same three knobs, not to argue about whether hand-written CoT is going extinct.