For most of the history of artificial intelligence, teaching a model to reason meant showing it examples of good reasoning. Researchers would write out step-by-step chains of thought — "first, calculate X, then check Y, then conclude Z" — and fine-tune the model to imitate those patterns. The approach works, but it also bakes in a ceiling: the model can only learn reasoning strategies that humans already know how to articulate and are willing to laboriously annotate.
The Ring-Zero project asks what happens when you remove that ceiling entirely. Instead of supervised fine-tuning on human examples, the team used reinforcement learning with verifiable rewards (RLVR), or "Zero RL" — an approach where the model explores reasoning paths freely and is rewarded only for arriving at correct answers, with no human telling it how to get there. The result, Ring-2.5-1T-Zero, is a 1-trillion-parameter Mixture-of-Experts model (63 billion activated parameters per query) that achieved an 84.2% accuracy on the AIME 2026 mathematics benchmark after its first RL stage, entirely from a pretrained base with no supervised reasoning data.
The key challenge of scaling Zero RL to a trillion parameters is instability. Very large models, when given free rein to explore, tend to generate excessively long and redundant reasoning chains — a phenomenon the researchers call "token redundancy." They also suffer from training instability as the policy gradient updates push the model away from its stable pretrained state. Ring-Zero addresses these problems through a four-stage training pipeline carefully designed to balance exploration, compression, and precision.
Stage one, reasoning elicitation, bootstraps the model by using token-level loss — calculating rewards based on the cumulative contribution of every token in a sequence — which explicitly encourages the model to generate longer, more detailed chains. A KL divergence penalty prevents it from drifting too far from its original parameters. Stage two, self-distillation and compression, takes the model's best reasoning traces, asks it to identify and remove redundant steps, and fine-tunes it on the compressed versions. This dramatically improves readability while preserving correctness. Stage three shifts to sample-level loss, normalizing gradients by response length so the model cannot game the reward system by simply being verbose; this forces it to sharpen its reasoning within tighter token budgets. Stage four introduces tier-based adaptive depth, training the model to allocate different amounts of reasoning effort depending on problem difficulty — solving simple arithmetic quickly while deploying deep chains for complex proofs.
Perhaps the most striking result is what the researchers observed during training: the model spontaneously developed cognitive strategies it was never explicitly taught. It learned to self-verify its intermediate steps, backtrack when it detected an error, and even explore multiple solution paths in parallel before converging on the best one. These behaviors were not programmed or prompted — they emerged naturally as the model scaled. This suggests that complex reasoning strategies may be latent capabilities that only require the right training regime and sufficient model capacity to surface.
Knowledge takeaway: Ring-Zero is a four-stage zero reinforcement learning pipeline that trains a 1-trillion-parameter MoE model to reason without any human-annotated reasoning data; the model achieved 84.2% on AIME 2026 math benchmark after the first RL stage; it spontaneously developed self-verification, backtracking, and parallel reasoning — strategies never explicitly programmed; the pipeline addresses training instability and token redundancy through a carefully sequenced protocol of elicitation, compression, length control, and adaptive depth allocation.