What's this about?
RLVR — reinforcement learning with verifiable rewards — powers today's reasoning models but only works where answers can be checked mechanically (math, code). This paper transforms open-ended tasks into forms where the model can verify its own rewards, extending the recipe to fuzzier domains.
What exactly is "RLVR"?
Reinforcement learning needs a reward signal. For math you can check the final answer; for code you can run the tests — "verifiable rewards." For essays or advice there's no checker, so those domains have been locked out of the technique. Making rewards *self*-verifiable is a bid to unlock them (18 upvotes).
Why it matters
Whether RL-style gains transfer beyond formal domains is one of the live debates (emollick posted about exactly this yesterday); this is a concrete attempt at the bridge.