An introduction to RL environments: how agents observe, act, and receive rewards, the shift from human grading to verifiable rewards (RLVR) and rubrics, and scaling agent training.

An introduction to RL environments, detailing verifiable rewards (RLVR), rubric-based evaluation, reward hacking, and how automated reward computation scales agent training.

What Are RL Environments? Rubrics, Verifiable Rewards, and How Agents Learn at Scale

A rubric is a named, structured list of criteria for judging a response or an action, with each criterion checkable on its own rather than folded into one overall score.

What exactly counts as a rubric?

It depends on scale. Domain experts write them by hand for high-stakes, narrow tasks. For everyday large-scale training, rubrics are usually generated synthetically or elicited automatically from pairwise human comparisons.

Who writes the rubrics used in real RL training?

Yes. Models can satisfy the letter of a criterion without performing the underlying task, a failure mode known as reward hacking. It is why verifiable rewards are preferred when available and why modern rubrics evolve during training.

Reinforcement Learning Cookbooks

What Are RL Environments? Rubrics, Verifiable Rewards, and Scaling