Backdoors in RLVR: Jailbreak Backdoors in LLMs From Verifiable Reward

Abstract

Reinforcement Learning with Verifiable Rewards (RLVR) is an emerging paradigm that significantly boosts a Large Language Model's (LLM's) reasoning abilities on complex logical tasks, such as mathematics and programming. However, we identify, for the first time, a latent vulnerability to backdoor attacks within the RLVR framework. This attack can implant a backdoor without modifying the reward verifier by injecting a small amount of poisoning data into the training set. Specifically, we propose a

Categories

Framework mappings

Suggested from the entry's categories.

@misc{guo2026backdoors,
  title = {{Backdoors in RLVR: Jailbreak Backdoors in LLMs From Verifiable Reward}},
  author = {Weiyang Guo and Zesheng Shi and Zeen Zhu and Yuan Zhou and Min Zhang and Jing Li},
  year = {2026},
  month = apr,
  eprint = {2604.09748},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.09748}
}