Unverified50% confidenceFactExact time
Traditional LLMs, after RLHF human preference alignment, tend to give confidently worded answers, which distorts probability estimates.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRLHF is intended to align the model with human preferences, but human preferences tend to reward 'confident and detailed' responses, creating a paradox that reinforces hallucination75% similarVerifiedIn the RLHF mechanism, human annotators tend to give higher scores to responses that exceed expectations and provide additional details, thereby systematically encouraging the model to perform inferential completion75% similarUnverifiedRLHF raters typically prefer responses that are fluently worded, measured in stance, and non-controversial, causing models to develop a lowest-common-denominator style of expression.74% similarUnverifiedIn RLHF, if preference data is of low quality or contains systematic bias, the reward model will learn incorrect signals, causing the model's behavior to deviate from expectations.74% similarUnverified主流LLM采用下一词预测目标，模型优化的是语言分布拟合度而非事实准确性，RLHF无法从根本上消除幻觉，完全消除幻觉在现有架构下被普遍认为不可能74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/943259API
curl https://kongchang.com/api/v1/knowledge/claims/943259MCP
get_claim(id=943259)