What's this about?
"Zero RL" — reinforcement learning with verifiable rewards and no human-annotated data — has mostly been studied on small models for cost reasons. Ring-Zero scales the recipe to a trillion-parameter model and reports emergent reasoning. Today's top HF Daily Papers pick (13 upvotes).
What exactly is "zero RL"?
Standard model training finishes with humans teaching the model what good answers look like (labels or preferences). Zero RL skips that: the model practices on problems where correctness is *automatically checkable* (math, code that runs) and learns purely from verified rewards — like learning chess by playing rather than by being shown grandmaster games. DeepSeek's R1-Zero made the idea famous; the question was whether it holds at giant scale.
Why is it worth noting?
It pairs with today's DeepSeek V4 recipe analysis (RLVR, no RLHF): the field's center of gravity is visibly shifting from human-preference training to verifiable-reward training.