RLVR

RLVR — reinforcement learning with verifiable rewards — is post-training in which the reward comes from a check the environment can run, not from a learned preference model or a human rating: a test suite passes, an answer matches, a program compiles. Because the reward is computed rather than judged, it cannot drift with an annotator’s taste — but it inherits every blind spot of the check itself, and the model optimises the check, not the intention behind it.

---

The term names the second stage of the now-standard two-stage recipe: SFT to put competent behaviour in reach, then RLVR to push into it. AfterQuery’s Terminal-Bench write-up is a clean worked example — the verifiable reward is a task’s own test suite run inside the container — and also a clean example of the blind-spot clause: their attempt to reward exploratory first turns directly produced exploration that was “performative rather than functional,” and the behaviour they wanted only emerged once the reward was again nothing but the tests passing.

The boundary worth keeping: RLVR names where the reward comes from, not which algorithm consumes it. AfterQuery drives it with [[grpo]]; anything that eats a scalar reward qualifies.