Verifiable Reward
A form of reward in reinforcement learning that can be automatically judged as correct or incorrect by a program. It applies to tasks such as mathematical proofs, code execution, and game outcomes, providing objective, low-noise training signals. It represents an important paradigm for reinforcement learning training of large language models.
Timeline (last 90 days)
RLHF's reward comes from human preferences, whereas the reward of RL post-training for control tasks comes from the environment itself (such as whether the task is completed, the magnitude of trajectory error), which is a verifiable reward.
The advantage of verifiable rewards is that the signal is objective and automatically computable; the drawback is that a simulation environment that can actually run is required.
Verifiable rewards refer to types of problems whose correctness can be automatically determined by a program, such as mathematical proofs, code execution results, and game outcomes
All Facts (3)
RLHF's reward comes from human preferences, whereas the reward of RL post-training for control tasks comes from the environment itself (such as whether the task is completed, the magnitude of trajectory error), which is a verifiable reward.
50%UnverifiedThe advantage of verifiable rewards is that the signal is objective and automatically computable; the drawback is that a simulation environment that can actually run is required.
50%UnverifiedVerifiable rewards refer to types of problems whose correctness can be automatically determined by a program, such as mathematical proofs, code execution results, and game outcomes
50%