[{"data":1,"prerenderedAt":838},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"wiki-temperature-top-p":9},{"tools":4,"reviews":5},65,7,{"tools":4,"reviews":5,"playbooks":7,"news":8},10,8,{"id":10,"title":11,"body":12,"category":820,"description":62,"extension":821,"meta":822,"navigation":823,"path":824,"published":825,"relatedModels":826,"relatedTools":829,"seo":830,"slug":831,"stem":832,"summary":833,"tags":834,"updated":825,"__hash__":837},"wiki\u002Fwiki\u002Ftemperature-top-p.md","Temperature 与 Top-P（采样参数）",{"type":13,"value":14,"toc":791},"minimark",[15,20,24,41,45,49,52,63,77,80,157,160,163,183,187,190,197,209,215,218,275,279,282,288,294,297,301,308,314,389,392,405,408,411,417,420,537,540,618,628,632,638,653,656,660,668,710,720,723,727,730,734,737,741,744,748,751,755,763,766,787],[16,17,19],"h2",{"id":18},"什么是-temperature-和-top-p","什么是 Temperature 和 Top-P",[21,22,23],"p",{},"Temperature 和 Top-P 是控制 LLM 输出随机性的两个参数。它们决定了模型在生成文本时\"有多保守\"或\"有多创造性\"。",[25,26,27,35],"ul",{},[28,29,30,34],"li",{},[31,32,33],"strong",{},"Temperature"," — 调整概率分布的平坦度，值越高输出越随机",[28,36,37,40],{},[31,38,39],{},"Top-P","（核采样） — 只从累积概率超过 P 的候选词中选，限制选择范围",[16,42,44],{"id":43},"temperature-详解","Temperature 详解",[46,47,48],"h3",{"id":48},"原理",[21,50,51],{},"模型在每一步预测下一个 token 时，会计算所有可能 token 的概率。Temperature 通过调整 logits（概率前的分数）来改变分布形状：",[53,54,59],"pre",{"className":55,"code":57,"language":58},[56],"language-text","调整后概率 = softmax(logits \u002F temperature)\n","text",[60,61,57],"code",{"__ignoreMap":62},"",[25,64,65,68,71,74],{},[28,66,67],{},"Temperature = 1.0：原始概率分布不变",[28,69,70],{},"Temperature \u003C 1.0：概率分布变\"尖锐\"——高概率词更高、低概率词更低 → 更确定",[28,72,73],{},"Temperature > 1.0：概率分布变\"平坦\"——各词概率更均匀 → 更随机",[28,75,76],{},"Temperature = 0：完全贪心——永远选概率最高的词",[46,78,79],{"id":79},"实际效果",[81,82,83,98],"table",{},[84,85,86],"thead",{},[87,88,89,92,95],"tr",{},[90,91,33],"th",{},[90,93,94],{},"效果",[90,96,97],{},"适用场景",[99,100,101,113,124,135,146],"tbody",{},[87,102,103,107,110],{},[104,105,106],"td",{},"0",[104,108,109],{},"完全确定性，每次回答相同",[104,111,112],{},"代码生成、数据抽取、事实问答",[87,114,115,118,121],{},[104,116,117],{},"0.3",[104,119,120],{},"高度确定，偶尔有变化",[104,122,123],{},"代码审查、文档摘要",[87,125,126,129,132],{},[104,127,128],{},"0.7",[104,130,131],{},"平衡（大多数 API 默认值）",[104,133,134],{},"通用对话、问答",[87,136,137,140,143],{},[104,138,139],{},"1.0",[104,141,142],{},"较有创造性",[104,144,145],{},"文案写作、头脑风暴",[87,147,148,151,154],{},[104,149,150],{},"1.5+",[104,152,153],{},"高度随机，可能出现乱码",[104,155,156],{},"创意写作（慎用）",[46,158,159],{"id":159},"示例",[21,161,162],{},"同一个 prompt \"写一首关于秋天的诗\"：",[25,164,165,171,177],{},[28,166,167,170],{},[31,168,169],{},"Temp 0","：每次生成完全相同的诗",[28,172,173,176],{},[31,174,175],{},"Temp 0.7","：每次不同的诗，但风格相似",[28,178,179,182],{},[31,180,181],{},"Temp 1.5","：每次差异巨大，可能出现非常规表达",[16,184,186],{"id":185},"top-p-详解","Top-P 详解",[46,188,48],{"id":189},"原理-1",[21,191,192,193,196],{},"Top-P（nucleus sampling，核采样）不是调整概率分布，而是",[31,194,195],{},"限制候选范围","：",[198,199,200,203,206],"ol",{},[28,201,202],{},"把所有候选 token 按概率从高到低排序",[28,204,205],{},"累积概率，直到达到 P 值",[28,207,208],{},"只从这些 token 中采样",[53,210,213],{"className":211,"code":212,"language":58},[56],"P = 0.9 → 只从累积概率达 90% 的最可能 token 中选\nP = 0.1 → 只选概率最高的极少数 token\nP = 1.0 → 不限制，所有 token 都可能被选\n",[60,214,212],{"__ignoreMap":62},[46,216,79],{"id":217},"实际效果-1",[81,219,220,230],{},[84,221,222],{},[87,223,224,226,228],{},[90,225,39],{},[90,227,94],{},[90,229,97],{},[99,231,232,243,254,265],{},[87,233,234,237,240],{},[104,235,236],{},"0.1",[104,238,239],{},"非常保守",[104,241,242],{},"代码生成、事实问答",[87,244,245,248,251],{},[104,246,247],{},"0.5",[104,249,250],{},"较保守",[104,252,253],{},"文档摘要、分类",[87,255,256,259,262],{},[104,257,258],{},"0.9",[104,260,261],{},"平衡（默认）",[104,263,264],{},"通用对话",[87,266,267,269,272],{},[104,268,139],{},[104,270,271],{},"不限制",[104,273,274],{},"创意写作",[16,276,278],{"id":277},"min_p新一代采样","min_p：新一代采样",[21,280,281],{},"2024 年后流行的第三个参数，部分开源推理框架（vLLM \u002F llama.cpp \u002F SGLang）和一些 API 已支持。",[21,283,284,287],{},[31,285,286],{},"思路","：top-p 在\"概率分布很尖\"时会留太多噪音 token；min_p 设一个相对阈值——只要 token 的概率不低于「最高概率 token × min_p」就保留。",[53,289,292],{"className":290,"code":291,"language":58},[56],"min_p = 0.05 表示：保留所有概率 ≥ 0.05 × max_prob 的 token\n",[60,293,291],{"__ignoreMap":62},[21,295,296],{},"实测在创意写作场景，min_p=0.05~0.1 比 top_p=0.9 输出质量更稳定（既不过度保守也不会跑飞）。新模型推理时可以试试。",[16,298,300],{"id":299},"seed-与确定性","seed 与确定性",[21,302,303,304,307],{},"Temperature=0 ",[31,305,306],{},"不等于"," 完全确定性。同一个 temperature=0 的请求，OpenAI \u002F Anthropic 多次跑结果有时仍不同——浮点运算非确定性 + batch 调度差异导致。",[21,309,310,311,196],{},"要更稳的复现，可以传 ",[60,312,313],{},"seed",[53,315,319],{"className":316,"code":317,"language":318,"meta":62,"style":62},"language-python shiki shiki-themes github-light github-dark","# OpenAI\nclient.chat.completions.create(\n    model=\"gpt-5\", temperature=0, seed=42, ...\n)\n# 响应里会带 system_fingerprint，相同 fingerprint + seed 才能保证完全一致\n","python",[60,320,321,330,337,377,383],{"__ignoreMap":62},[322,323,326],"span",{"class":324,"line":325},"line",1,[322,327,329],{"class":328},"sJ8bj","# OpenAI\n",[322,331,333],{"class":324,"line":332},2,[322,334,336],{"class":335},"sVt8B","client.chat.completions.create(\n",[322,338,340,344,348,352,355,358,360,363,365,367,369,372,374],{"class":324,"line":339},3,[322,341,343],{"class":342},"s4XuR","    model",[322,345,347],{"class":346},"szBVR","=",[322,349,351],{"class":350},"sZZnC","\"gpt-5\"",[322,353,354],{"class":335},", ",[322,356,357],{"class":342},"temperature",[322,359,347],{"class":346},[322,361,106],{"class":362},"sj4cs",[322,364,354],{"class":335},[322,366,313],{"class":342},[322,368,347],{"class":346},[322,370,371],{"class":362},"42",[322,373,354],{"class":335},[322,375,376],{"class":362},"...\n",[322,378,380],{"class":324,"line":379},4,[322,381,382],{"class":335},")\n",[322,384,386],{"class":324,"line":385},5,[322,387,388],{"class":328},"# 响应里会带 system_fingerprint，相同 fingerprint + seed 才能保证完全一致\n",[21,390,391],{},"注意：",[25,393,394,399,402],{},[28,395,396,398],{},[60,397,313],{}," 是 best-effort，模型升级 \u002F 基础设施变动会让 fingerprint 变",[28,400,401],{},"Anthropic \u002F Google 早期不支持 seed，新版本逐步加入",[28,403,404],{},"真的要 100% 确定性（比如单元测试），用 mock 替代 LLM 调用",[16,406,407],{"id":407},"怎么搭配使用",[46,409,410],{"id":410},"一般原则",[21,412,413,416],{},[31,414,415],{},"不要同时调两个","。OpenAI 官方建议：要么调 Temperature，要么调 Top-P，不要同时改。",[46,418,419],{"id":419},"推荐配置",[81,421,422,436],{},[84,423,424],{},[87,425,426,429,431,433],{},[90,427,428],{},"场景",[90,430,33],{},[90,432,39],{},[90,434,435],{},"理由",[99,437,438,451,464,476,489,501,512,524],{},[87,439,440,443,445,448],{},[104,441,442],{},"代码生成",[104,444,106],{},[104,446,447],{},"1",[104,449,450],{},"完全确定，代码不应有\"创意\"",[87,452,453,456,459,461],{},[104,454,455],{},"代码审查",[104,457,458],{},"0.2",[104,460,447],{},[104,462,463],{},"高度确定，偶尔看不同角度",[87,465,466,469,471,473],{},[104,467,468],{},"数据抽取",[104,470,106],{},[104,472,447],{},[104,474,475],{},"严格按格式输出",[87,477,478,481,484,486],{},[104,479,480],{},"工具调用（Function Calling）",[104,482,483],{},"0 ~ 0.2",[104,485,447],{},[104,487,488],{},"见下节",[87,490,491,494,496,498],{},[104,492,493],{},"客服 Bot",[104,495,247],{},[104,497,258],{},[104,499,500],{},"适度变化，但不跑题",[87,502,503,505,507,509],{},[104,504,264],{},[104,506,128],{},[104,508,447],{},[104,510,511],{},"平衡",[87,513,514,517,519,521],{},[104,515,516],{},"文案写作",[104,518,258],{},[104,520,447],{},[104,522,523],{},"需要创意",[87,525,526,529,532,534],{},[104,527,528],{},"头脑风暴",[104,530,531],{},"1.2",[104,533,447],{},[104,535,536],{},"越发散越好",[46,538,539],{"id":539},"各平台默认值",[81,541,542,558],{},[84,543,544],{},[87,545,546,549,552,555],{},[90,547,548],{},"平台",[90,550,551],{},"默认 Temperature",[90,553,554],{},"默认 Top-P",[90,556,557],{},"备注",[99,559,560,572,583,595,607],{},[87,561,562,565,567,569],{},[104,563,564],{},"OpenAI API",[104,566,139],{},[104,568,139],{},[104,570,571],{},"—",[87,573,574,577,579,581],{},[104,575,576],{},"Anthropic API",[104,578,139],{},[104,580,139],{},[104,582,571],{},[87,584,585,588,590,593],{},[104,586,587],{},"Google Gemini",[104,589,139],{},[104,591,592],{},"0.95",[104,594,571],{},[87,596,597,600,602,604],{},[104,598,599],{},"DeepSeek API",[104,601,139],{},[104,603,139],{},[104,605,606],{},"官方文档建议代码 0.0 \u002F 通用 1.3",[87,608,609,612,614,616],{},[104,610,611],{},"国内 GLM",[104,613,592],{},[104,615,128],{},[104,617,571],{},[21,619,620,623,624,627],{},[31,621,622],{},"踩坑","：很多人以为默认是 0.7，其实主流大厂默认是 1.0。如果你的应用想要保守输出，",[31,625,626],{},"必须显式设置","，不要假设默认值。",[16,629,631],{"id":630},"推理模型为什么不建议改-temperature","推理模型为什么不建议改 temperature",[21,633,634,635,196],{},"GPT-5 reasoning、Claude Opus 4 thinking、DeepSeek-R1、o3 等推理模型，官方都建议",[31,636,637],{},"保持默认 temperature=1.0",[25,639,640,643,650],{},[28,641,642],{},"推理模型的思维链本身依赖概率采样的多样性来\"探索\"不同解法",[28,644,645,646,649],{},"强制 t=0 会让推理路径单一，反而",[31,647,648],{},"降低","复杂问题的成功率",[28,651,652],{},"思维链长度 × 低多样性 = 容易陷入死循环",[21,654,655],{},"对推理模型，控制输出应该靠 prompt 描述目标和约束，而不是调采样参数。",[16,657,659],{"id":658},"温度对-function-calling-稳定性的影响","温度对 Function Calling 稳定性的影响",[21,661,662,667],{},[663,664,666],"a",{"href":665},"\u002Fwiki\u002Ffunction-calling.html","工具调用","场景，temperature 影响很关键：",[81,669,670,679],{},[84,671,672],{},[87,673,674,676],{},[90,675,33],{},[90,677,678],{},"影响",[99,680,681,688,695,702],{},[87,682,683,685],{},[104,684,106],{},[104,686,687],{},"工具选择最稳定，参数提取准确率最高",[87,689,690,692],{},[104,691,458],{},[104,693,694],{},"几乎和 0 一样，偶尔在多个相近工具间选",[87,696,697,699],{},[104,698,128],{},[104,700,701],{},"模型可能\"灵活\"地选不太对的工具、参数飘",[87,703,704,707],{},[104,705,706],{},"1.0+",[104,708,709],{},"工具调用稳定性显著下降，生产慎用",[21,711,712,715,716,719],{},[31,713,714],{},"经验","：所有工具调用场景，先把 temperature 设 0 验证基线效果，再视情况微调。",[31,717,718],{},"别用默认 1.0 直接上工具调用","，调试起来很痛苦。",[16,721,722],{"id":722},"常见误区",[46,724,726],{"id":725},"误区-1-temperature-0-就不会出错","误区 1: \"Temperature = 0 就不会出错\"",[21,728,729],{},"Temperature = 0 只保证（接近）确定性，不保证正确性。模型可能每次都\"确定地\"给出错误答案。",[46,731,733],{"id":732},"误区-2-调高-temperature-模型更聪明","误区 2: \"调高 Temperature 模型更聪明\"",[21,735,736],{},"Temperature 调高只是让输出更随机，不会让模型变聪明。高 Temperature 反而可能导致逻辑混乱。",[46,738,740],{"id":739},"误区-3-top-p-和-temperature-效果一样","误区 3: \"Top-P 和 Temperature 效果一样\"",[21,742,743],{},"不完全一样。Top-P 是硬截断（低概率词完全排除），Temperature 是软调整（低概率词概率降低但不为 0）。",[46,745,747],{"id":746},"误区-4-推理模型也要-temp0","误区 4: \"推理模型也要 temp=0\"",[21,749,750],{},"错。推理模型靠思维链探索，强制 0 会让它失去推理多样性，反而表现下降。",[46,752,754],{"id":753},"误区-5-改-temperature-可以解决幻觉","误区 5: \"改 temperature 可以解决幻觉\"",[21,756,757,758,762],{},"只能轻微缓解。",[663,759,761],{"href":760},"\u002Fwiki\u002Fhallucination.html","幻觉","的根本解法是 RAG + grounded generation，不是调采样参数。",[16,764,765],{"id":765},"延伸阅读",[25,767,768,775,781],{},[28,769,770,771],{},"输出控制：",[663,772,774],{"href":773},"\u002Fwiki\u002Fprompt-engineering.html","Prompt Engineering",[28,776,777,778],{},"工具稳定性：",[663,779,780],{"href":665},"Function Calling",[28,782,783,784],{},"幻觉缓解：",[663,785,786],{"href":760},"Hallucination",[788,789,790],"style",{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":62,"searchDepth":339,"depth":339,"links":792},[793,794,799,803,804,805,810,811,812,819],{"id":18,"depth":332,"text":19},{"id":43,"depth":332,"text":44,"children":795},[796,797,798],{"id":48,"depth":339,"text":48},{"id":79,"depth":339,"text":79},{"id":159,"depth":339,"text":159},{"id":185,"depth":332,"text":186,"children":800},[801,802],{"id":189,"depth":339,"text":48},{"id":217,"depth":339,"text":79},{"id":277,"depth":332,"text":278},{"id":299,"depth":332,"text":300},{"id":407,"depth":332,"text":407,"children":806},[807,808,809],{"id":410,"depth":339,"text":410},{"id":419,"depth":339,"text":419},{"id":539,"depth":339,"text":539},{"id":630,"depth":332,"text":631},{"id":658,"depth":332,"text":659},{"id":722,"depth":332,"text":722,"children":813},[814,815,816,817,818],{"id":725,"depth":339,"text":726},{"id":732,"depth":339,"text":733},{"id":739,"depth":339,"text":740},{"id":746,"depth":339,"text":747},{"id":753,"depth":339,"text":754},{"id":765,"depth":332,"text":765},"concept","md",{},true,"\u002Fwiki\u002Ftemperature-top-p","2026-06-21",[827,828],"claude-sonnet-4","gpt-5",null,{"title":11,"description":62},"temperature-top-p","wiki\u002Ftemperature-top-p","控制 LLM 输出随机性的两个核心参数：Temperature 调节概率分布的平坦度，Top-P 限制候选词范围。",[33,39,835,836],"采样","参数调优","BvXFYUS66mHKafqko--Hxl2TOXnOvIwrSx9s9ZhMrBY",1782316491081]