Unverified50% confidenceFactExact time
The Process Reward Model (PRM) applies the reward signal step by step to each reasoning link, which can alleviate the reward hacking risk caused by traditional RLHF giving rewards only at the final answer.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedProcess Reward Models (PRM) shift toward evaluating the reasoning process rather than the final answer, and have already demonstrated significant advantages in the domain of mathematical reasoning.77% similarUnverifiedRLHF是InstructGPT和ChatGPT能力跃升的核心技术，但存在奖励模型被过度优化（reward hacking）的内在风险73% similarVerifiedRLHF的核心机制是训练一个奖励模型来预测人类对输出的偏好评分，再以奖励信号驱动策略梯度优化（通常采用PPO算法）73% similarUnverifiedThe reward in that experiment is provided by a PRM (process reward model), which is a frozen base version of the 4B policy model, and the entire process runs on seven GPUs.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1028458API
curl https://kongchang.com/api/v1/knowledge/claims/1028458MCP
get_claim(id=1028458)