Unverified50% confidenceFactExact time
在RLHF训练阶段模型被持续奖励产生'有帮助的'输出，导致其倾向于提供超出请求范围的内容，即'过度生成'（over-generation）
1
Sources
50%
Confidence
Long-term
Relevance
8/25/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedRLHF trains a reward model by collecting human preference ratings, then fine-tunes the language model using reinforcement learning algorithms such as PPO; however, the reward model can be deceived by adversarial inputs and cannot enumerate all variants of harmful request expressions.74% similarUnverified提示词与模型版本存在深度绑定关系，不同版本模型在RLHF训练数据、指令微调策略上的差异会导致相同提示词产生不同输出，这种现象被称为「提示词漂移」74% similarUnverifiedModels being 'overly eager' is a common failure mode of AI coding agents, stemming from large language models being optimized during RLHF training to be 'as helpful as possible,' causing them to favor taking action over remaining passive.74% similarVerified大语言模型在RLHF训练阶段，因人类标注员倾向于给更详细、更礼貌的回答打高分，导致模型学会了过度冗长的表达模式74% similarUnverifiedThe model's tendency to stop stems from the RLHF (Reinforcement Learning from Human Feedback) training mechanism, where a large number of examples in the training data end the conversation after completing a single task.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/797088API
curl https://kongchang.com/api/v1/knowledge/claims/797088MCP
get_claim(id=797088)