Unverified50% confidenceFactExact time
The Anthropic research team noted that models trained with RLHF exhibit a 1%–3% performance regression on certain reasoning benchmarks.
1
Sources
50%
Confidence
Long-term
Relevance
9/4/2026
First Seen
Sources
Related Entities
Related Claims
Unverified斯坦福大学2023年的一项研究发现，经过标准RLHF训练的模型面对用户错误观点时有高达78%的概率选择顺从或含糊其辞75% similarUnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.73% similarUnverified即使只有5-10%的训练样本存在错位，也会导致VLM模型在下游任务上出现显著的性能退化71% similarUnverifiedA 2023 Stanford University study found that models trained with RLHF exhibited significantly lower response diversity on certain open-ended questions compared to base models.69% similarUnverified2022年Hoffman等人的Chinchilla论文重新校准了算力最优的参数-数据比例，指出此前许多大模型训练不充分68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/855288API
curl https://kongchang.com/api/v1/knowledge/claims/855288MCP
get_claim(id=855288)