Expired50% confidenceFactTime unknown
Research data indicates that when generated text exceeds 2,000 tokens, the hallucination rate may be 2–3 times higher than in short responses.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026(expired)
Sources
Related Claims
Unverified有研究表明当输入文本具有明确结构化标记时，主流LLM的信息提取准确率可提升20%以上，幻觉发生率相应降低71% similarVerifiedResearch shows that even the most advanced models have hallucination rates in factual question-answering tasks that still fluctuate between 5% and 20%.67% similarUnverified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63% similarUnverified研究表明在要求AI直接生成参考文献的任务中，幻觉率可高达40%-70%63% similarUnverified超高思考强度一条消息消耗的token可能是低强度的5-10倍63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/55872API
curl https://kongchang.com/api/v1/knowledge/claims/55872MCP
get_claim(id=55872)