Unverified50% confidenceFactExact time
在专业信息上对模型进行微调可能导致更多幻觉，且成本效益极差
1
Sources
50%
Confidence
Long-term
Relevance
7/22/2026
First Seen
Sources
Claude Code高阶实战：打造顶级AI开发流
Related Claims
Unverified早期模型在某些领域的事实幻觉率高达15-25%，尤其在长尾知识领域幻觉率更高73% similarVerifiedResearch shows that even the most advanced models have hallucination rates in factual question-answering tasks that still fluctuate between 5% and 20%.72% similarUnverified研究显示，在专业技术领域，LLM 的幻觉率显著高于通用知识领域，因为训练数据中领域专家的纠错反馈相对稀少71% similarUnverifiedWhen the context exceeds 40% of the model's context window, the model's reasoning ability noticeably degrades, exhibiting a decline in intelligence.70% similarUnverified斯坦福大学研究发现，在推理链关键步骤注入错误信息对最终答案的影响远小于预期69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/587662API
curl https://kongchang.com/api/v1/knowledge/claims/587662MCP
get_claim(id=587662)