Verifiable Rewards
also called RLVR, RL from verifiable rewards
Training against an answer a machine can check.
Reinforcement learning where the reward comes from an automatic check rather than human preference — a unit test passing, a maths answer matching, a proof validating. Because the signal is objective, it scales to far more training examples than human labelling.
Its limit is the shape of the tasks: it works where correctness is machine-checkable, which is maths, code and structured problems. That is exactly why the reasoning-model wave arrived first in those domains, and why domains without a checker lag behind.
The clean way to think about it: a reward you cannot argue with is also a reward the model can game if it can reach the checker. Verifiable rewards make training tractable and make reward hacking concrete, which is why the same body of work contains both the capability gains and the cautionary findings.
Building your own checkable signal
A team wants better performance on an internal domain.
If correctness is checkable — a reconciliation balances, a query returns the expected rows, a config validates — you can generate many training examples automatically. Where it is not checkable, you are back to human preference data and its costs.
- paperDeepSeek-R1: incentivizing reasoning capability in LLMs via reinforcement learning ↗
Reinforcement learning on verifiable tasks elicited reasoning patterns — self-checking, backtracking — with substantial open publication of the recipe.
- paperEvaluating large language models trained on code ↗
HumanEval, and the formal pass@k estimator: the chance that at least one of k samples passes the tests. Not the same as passing k times in a row.
RLHF · Reinforcement Learning from Human Feedback
Tuning a model toward what humans rate as better answers.
Reasoning Model · Thinking model
A model trained to spend tokens thinking before it answers.
Reward Hacking · Goodharting
Optimising the metric instead of the goal.
Fine-tuning
Extra training on your own examples to change how the model behaves.
Automated Check
A machine verdict the agent cannot talk its way past.