<!DOCTYPE html>
<html lang="zh-CN">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>人造蜂巢意识:语言模型中的开放式同质性现象</title>
<link href="https://fonts.googleapis.com/icon?family=Material+Icons" rel="stylesheet">
<link href="https://fonts.googleapis.com/css2?family=Noto+Sans+SC:wght@400;500;700;900&display=swap" rel="stylesheet">
<style>
* {
margin: 0;
padding: 0;
box-sizing: border-box;
}
body {
font-family: 'Noto Sans SC', sans-serif;
background-color: #f5f5f5;
color: #333;
line-height: 1.6;
}
.poster-container {
width: 720px;
min-height: 960px;
margin: 0 auto;
background: linear-gradient(135deg, #0a1929 0%, #132f4c 100%);
position: relative;
overflow: hidden;
padding: 40px 30px;
color: #fff;
}
/* 六边形背景图案 */
.hexagon-bg {
position: absolute;
top: 0;
left: 0;
width: 100%;
height: 100%;
opacity: 0.05;
background-image: url('https://sfile.chatglm.cn/moeSlide/image/92/92938cb1.jpg');
background-size: cover;
z-index: 0;
}
.content {
position: relative;
z-index: 1;
}
/* 标题区域 */
.header {
text-align: center;
margin-bottom: 30px;
padding: 20px;
border-radius: 15px;
background: rgba(0, 0, 0, 0.2);
backdrop-filter: blur(10px);
border: 1px solid rgba(255, 255, 255, 0.1);
}
.main-title {
font-size: 40px;
font-weight: 900;
margin-bottom: 10px;
color: #fff;
line-height: 1.3;
text-shadow: 0 2px 4px rgba(0, 0, 0, 0.3);
}
.sub-title {
font-size: 24px;
font-weight: 500;
margin-bottom: 15px;
color: #e0e0e0;
}
.award-badge {
display: inline-block;
background: linear-gradient(45deg, #ff7e5f, #feb47b);
color: white;
padding: 8px 20px;
border-radius: 30px;
font-weight: 700;
font-size: 20px;
box-shadow: 0 4px 15px rgba(255, 126, 95, 0.4);
transform: rotate(-2deg);
margin-top: 10px;
}
/* 内容卡片 */
.card {
background: rgba(255, 255, 255, 0.08);
backdrop-filter: blur(10px);
border-radius: 15px;
padding: 25px;
margin-bottom: 25px;
border: 1px solid rgba(255, 255, 255, 0.1);
box-shadow: 0 8px 32px rgba(0, 0, 0, 0.1);
transition: transform 0.3s ease, box-shadow 0.3s ease;
}
.card:hover {
transform: translateY(-5px);
box-shadow: 0 12px 40px rgba(0, 0, 0, 0.2);
}
.card-title {
display: flex;
align-items: center;
margin-bottom: 15px;
color: #4fc3f7;
font-size: 28px;
font-weight: 700;
}
.card-title .material-icons {
margin-right: 10px;
font-size: 32px;
}
.card-content {
font-size: 18px;
line-height: 1.6;
}
.highlight {
color: #ffb74d;
font-weight: 700;
}
/* 研究发现区域 */
.findings {
display: flex;
justify-content: space-between;
margin-top: 20px;
}
.finding-item {
flex: 1;
background: rgba(0, 0, 0, 0.2);
padding: 15px;
border-radius: 10px;
margin: 0 5px;
text-align: center;
}
.finding-item:first-child {
margin-left: 0;
}
.finding-item:last-child {
margin-right: 0;
}
.finding-icon {
font-size: 36px;
color: #ffb74d;
margin-bottom: 10px;
}
.finding-title {
font-weight: 700;
margin-bottom: 8px;
color: #4fc3f7;
}
/* 图表区域 */
.chart-container {
background: rgba(0, 0, 0, 0.2);
border-radius: 10px;
padding: 15px;
margin-top: 20px;
text-align: center;
}
.chart-img {
max-width: 100%;
border-radius: 8px;
margin: 10px 0;
}
/* 底部信息 */
.footer {
margin-top: 30px;
text-align: center;
font-size: 16px;
color: #b0bec5;
padding: 20px;
border-radius: 15px;
background: rgba(0, 0, 0, 0.2);
backdrop-filter: blur(10px);
border: 1px solid rgba(255, 255, 255, 0.1);
}
.authors {
margin-bottom: 10px;
}
.institutions {
margin-bottom: 10px;
}
/* 装饰元素 */
.hex-decoration {
position: absolute;
width: 100px;
height: 100px;
opacity: 0.1;
z-index: 0;
}
.hex-1 {
top: 10%;
right: 5%;
transform: rotate(15deg);
}
.hex-2 {
bottom: 15%;
left: 5%;
transform: rotate(-10deg);
}
/* 数据点动画 */
<span class="mention-invalid">@keyframes</span> pulse {
0% {
transform: scale(1);
opacity: 0.8;
}
50% {
transform: scale(1.2);
opacity: 1;
}
100% {
transform: scale(1);
opacity: 0.8;
}
}
.data-point {
display: inline-block;
width: 12px;
height: 12px;
border-radius: 50%;
margin: 0 3px;
animation: pulse 2s infinite;
}
.data-point-1 {
background-color: #4fc3f7;
animation-delay: 0s;
}
.data-point-2 {
background-color: #ffb74d;
animation-delay: 0.5s;
}
.data-point-3 {
background-color: #ff8a65;
animation-delay: 1s;
}
.data-point-4 {
background-color: #81c784;
animation-delay: 1.5s;
}
/* 强调文本 */
.emphasis-box {
background: rgba(255, 183, 77, 0.15);
border-left: 4px solid #ffb74d;
padding: 15px;
margin: 15px 0;
border-radius: 0 8px 8px 0;
}
.emphasis-text {
font-size: 20px;
font-weight: 500;
}
</style>
</head>
<body>
<div class="poster-container">
<div class="hexagon-bg"></div>
<!-- 装饰六边形 -->
<div class="hex-decoration hex-1">
<i class="material-icons" style="font-size: 100px; color: #4fc3f7;">hexagon</i>
</div>
<div class="hex-decoration hex-2">
<i class="material-icons" style="font-size: 100px; color: #ffb74d;">hexagon</i>
</div>
<div class="content">
<!-- 标题区域 -->
<div class="header">
<h1 class="main-title">人造蜂巢意识:语言模型中的开放式同质性现象</h1>
<h2 class="sub-title">Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)</h2>
<div class="award-badge">NeurIPS 2025 最佳论文</div>
</div>
<!-- 研究背景 -->
<div class="card">
<h3 class="card-title">
<i class="material-icons">psychology</i>
研究背景
</h3>
<div class="card-content">
大语言模型在生成多样化、具有人类风格的创意内容方面表现欠佳,生成内容往往趋向于同质化。这种现象引发了人们对人类思维因反复接触相似输出而长期趋于同质化的担忧。
<div class="emphasis-box">
<p class="emphasis-text">所有 AI 模型都在趋同,形成了一种缺乏个性、千篇一律的"平庸集体共识",就像蜂群中的工蜂一样没有独立思想。</p>
</div>
</div>
</div>
<!-- 研究方法 -->
<div class="card">
<h3 class="card-title">
<i class="material-icons">science</i>
研究方法
</h3>
<div class="card-content">
<ul>
<li><span class="highlight">Infinity-Chat数据集</span>:包含26K条多样、真实世界、开放式的用户查询</li>
<li><span class="highlight">分类体系</span>:首个用于刻画语言模型所面对开放式提示全谱系的综合分类体系,包含6个顶层类别(创意内容生成、头脑风暴与构思等)和17个子类别</li>
<li><span class="highlight">人工标注</span>:31,250条人工标注,覆盖绝对评分与两两偏好选择,每个样本均有25位独立标注者参与</li>
</ul>
<div class="chart-container">
<img src="https://sfile.chatglm.cn/moeSlide/image/2b/2b9a33d1.jpg" alt="六边形数据结构" class="chart-img">
<p>Infinity-Chat数据集结构化组织示意图</p>
</div>
</div>
</div>
<!-- 主要发现 -->
<div class="card">
<h3 class="card-title">
<i class="material-icons">lightbulb</i>
主要发现
</h3>
<div class="card-content">
<p>研究揭示了在开放式生成中显著存在的<span class="highlight">人工蜂群思维效应</span>,其特征包括:</p>
<div class="findings">
<div class="finding-item">
<i class="material-icons finding-icon">repeat</i>
<div class="finding-title">模型内重复</div>
<p>同一个模型会反复生成相似的回答</p>
</div>
<div class="finding-item">
<i class="material-icons finding-icon">device_hub</i>
<div class="finding-title">模型间同质化</div>
<p>不同模型之间也会产生惊人相似的输出</p>
</div>
</div>
<div class="chart-container">
<img src="https://sfile.chatglm.cn/moeSlide/image/73/73862a2a.jpg" alt="模型同质性可视化" class="chart-img">
<p>模型间同质性现象可视化(不同颜色代表不同模型)</p>
<p>即便采用增强多样性的解码策略,仍有超60%的响应相似度超过0.8</p>
</div>
</div>
</div>
<!-- 研究意义 -->
<div class="card">
<h3 class="card-title">
<i class="material-icons">insights</i>
研究意义
</h3>
<div class="card-content">
<ul>
<li>为数据集与基准树立了新标杆</li>
<li>推动了对科学理解与重大社会挑战的进展,而不只是单纯提升技术性能</li>
<li>为缓解人工蜂群思维带来的长期AI安全风险提供了关键洞见</li>
<li>为未来开发更具多样性、更贴合人类多元需求的AI系统提供了基准和方向</li>
</ul>
<div class="emphasis-box">
<p class="emphasis-text">这项工作在理解现代语言模型中的多样性、价值多元与社会影响方面做出了重要且及时的贡献。</p>
</div>
</div>
</div>
<!-- 底部信息 -->
<div class="footer">
<div class="authors">
<strong>作者:</strong>Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, Yejin Choi
</div>
<div class="institutions">
<strong>机构:</strong>华盛顿大学、卡内基梅隆大学、艾伦人工智能研究所、Lila Sciences、斯坦福大学
</div>
<div class="conference">
<strong>会议:</strong>NeurIPS 2025
</div>
<div style="margin-top: 15px;">
<span class="data-point data-point-1"></span>
<span class="data-point data-point-2"></span>
<span class="data-point data-point-3"></span>
<span class="data-point data-point-4"></span>
</div>
</div>
</div>
</div>
</body>
</html>
登录后可参与表态
讨论回复
0 条回复还没有人回复,快来发表你的看法吧!