protologue

Reinforcement Learning with Verifiable Rewards

Also called RLVR, RL from verifiable rewards, outcome-based reinforcement learning.

Reinforcement learning with verifiable rewards (RLVR) trains a language model with rewards computed by an automatic check of correctness, such as whether a math answer matches the reference or code passes its tests, instead of by a learned reward model.

Description

Verifiable rewards are hard to game in the ways a learned reward model can be, and large-scale RL with them is how reasoning models learn to produce long, self-correcting chains of thought. The approach is limited to tasks with checkable answers, so open-ended writing and judgment still rely on preference-based rewards.

Sources

  1. Lambert et al. (2024). Tulu 3: Pushing Frontiers in Open Language Model Post-Training.
  2. DeepSeek-AI et al. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.

Cite this entry

Protologue. (2026). Reinforcement Learning with Verifiable Rewards. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0156). https://protologue.com/t/rlvr/

BibTeX
@misc{protologue_rlvr,
  title = {Reinforcement Learning with Verifiable Rewards},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0156},
  url = {https://protologue.com/t/rlvr/}
}

Markdown JSON