Binary Retrieval-Augmented Reward Mitigates Hallucinations (arxiv.org)

Researchers propose an online reinforcement-learning method called binary retrieval-augmented reward (RAR) to curb extrinsic hallucinations from language models. Instead of a continuous scalar, RAR gives a reward of 1 only when a model’s output is fully supported by retrieved evidence and factually correct, and 0 otherwise. Applied to Qwen3 reasoning models across tasks, this approach cut hallucination rates in open-ended generation by 39.3% versus supervised and continuous-reward RL baselines. In short-form QA the models learned calibrated abstention—responding “I don’t know” when evidence was lacking—reducing incorrect answers by 44.4% on PopQA and 21.7% on GPQA. Code, math, and instruction-following performance were preserved, whereas continuous-reward RL, while improving factuality, caused quality regressions. The key technical insight is combining retrieval-based factual verification with a binary, sparse RL signal that incentivizes wholly correct outputs rather than partial improvements, which appears to avoid the tradeoff between factuality and generation quality. Practically, this suggests a promising path for production LLMs to be both more truthful and maintain broad capabilities, though success will depend on reliable retrieval and verification pipelines and strategies to handle sparse rewards during training.