Does scale alone teach models to reason without hand-crafted rewards?

Ring-Zero scales "zero RL" — RLVR starting directly from a pretrained base with no supervised CoT data — to a trillion parameters, and reports a concrete scale threshold for the bitter lesson. On their 104B model they had to introduce a hand-designed reward to coax structured formatting, self-verification, and parallel-path exploration. On the 1T model those same heuristics became "completely redundant": under pure zero RL the model spontaneously discovered structured reasoning, self-verification, and other advanced cognitive strategies from scratch, with no auxiliary reward.

This is more than a bigger-is-better result. It says the engineering effort itself — the crafted rewards, curated SFT sets, and format-forcing prompts the field has invested in — is a scale-dependent crutch. At small scale the model cannot find these behaviors alone, so we inject them; at large scale the same injection is not just unnecessary but a heuristic the bitter lesson predicts general computation will overtake. Which means the value of much reasoning-scaffolding research has a shelf life tied to model size.

There is a tension worth flagging. The naive scaling path suffered poor readability, token redundancy, and no adaptive reasoning depth — the same Does more thinking time always improve reasoning accuracy? pathology — and Ring-Zero needed system-level fixes (clipped importance sampling, train-inference ratio correction, mixed-precision control) to stabilize training. So "scale removes the need for heuristics" holds for the reward heuristics while shifting the burden to systems engineering. The crutch moves rather than disappears.