Unverified50% confidenceTradeoffExact time
--stylize数值越低生成结果越忠实于提示词字面描述，数值越高模型引入更多美学判断但会偏离提示词精确指令
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Partially Verified接受率越高意味着草稿模型猜测越准，大模型返工次数越少，直接决定投机解码的实际加速效果76% similarUnverified提示词越模糊，模型输出越向高概率区域坍缩，一致性越强75% similarVerifiedTemperature参数越低越倾向选择概率最高的词输出更保守，越高输出更具创意但更易出现幻觉75% similarUnverified温度值越低模型输出越确定性，适合事实查询和代码生成；值越高输出更具创造性，适合文案创作75% similarUnverified在提示中设定量化约束（如对白比例、单句长度、段落密度）比模糊的风格描述更容易被模型稳定遵守74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/910170API
curl https://kongchang.com/api/v1/knowledge/claims/910170MCP
get_claim(id=910170)