Ant Group Scales Zero-RL to 1T Params in Ring-Zero Paper

Most 'zero-annotation reinforcement learning' work up to now has been done at relatively small model sizes, because throwing a trillion parameters at a reward-model-free training loop is expensive and, until recently, no one had shown it was worth it. A new arXiv paper from Ant Group's Ring team reports doing exactly that, training Ring-2.5-1T-Zero using reinforcement learning with verifiable rewards and no human-annotated data, and documenting what falls out at that scale.

The claim I find most interesting is not the parameter count. It is the list of behaviors the team says the model developed on its own during RL, without anyone hand-crafting heuristics for them: self-verification, parallel reasoning, structured formatting, plus what the authors call anthropomorphism and 'context anxiety'. Their reading of the training curve is that the model goes through two distinct stages, an initial discovery phase followed by a sharpening phase, and that these cognitive habits emerge across that arc.

To make trillion-scale zero RL stable at all, the researchers describe several plumbing fixes, clipped importance sampling, training-inference ratio correction, and mixed-precision control, aimed at problems like poor readability and token redundancy that they say naive scaling produces. Ring-2.5-1T-Zero is evaluated on seven mathematical benchmarks with what the paper calls competitive performance, and the team also introduces a separate evaluation framework for chain-of-thought quality along three axes, comprehensibility, reproducibility, and efficiency.

The honest caveat is that the retrieved abstract advertises 'competitive' scores without naming the seven benchmarks or the models being compared against, and says nothing about total training compute, token count, or base-model provenance. Those omissions matter, because a trillion-parameter result that is only competitive rather than dominant, at unknown cost, is a very different signal from a clean win. Take the emergent-behavior list as reported observations, not settled science.

If the pattern generalizes beyond math, the practical read is that teams working in domains where correctness is machine-checkable, math, code, formal proofs, verifiable tool use, can spend on compute and reward verifiers instead of on annotation pipelines. That is a meaningful reshuffle of where the money goes in a post-training program, and worth watching to see whether other labs reproduce it at this scale.