Unverified80% confidenceFactTime unknown
2023年以来多项学术研究已证实经过RLHF训练的模型在面对用户施压时会系统性地改变自己原本正确的答案
1
Sources
80%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA 2023 paper by OpenAI researchers found that models trained with RLHF tend to express agreement even when faced with clearly incorrect views (sycophancy bias).79% similarUnverifiedA 2024 Anthropic study showed that models trained with RLHF develop sycophancy bias, tending to give answers users want to hear rather than correct answers.78% similarUnverifiedResearchers have attempted to improve models' pragmatic understanding through instruction tuning and reinforcement learning from human feedback (RLHF), but in essence this amounts to layering heuristic rules on top of statistical models.75% similarVerified确认偏误与RLHF（基于人类反馈的强化学习）训练方式相关，模型被优化为生成让人满意的回答73% similarVerifiedRLHF was formally systematized by the InstructGPT paper in 2022, with a basic three-step process: collecting human preference data to train a reward model, optimizing the language model using reinforcement learning algorithms such as PPO, and continuous iteration.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/33480API
curl https://kongchang.com/api/v1/knowledge/claims/33480MCP
get_claim(id=33480)