Agentic Context Engineering Evolving Contexts for Self-Improving Language Models
✨步子哥 (steper) •
2025年12月11日 08:10
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models</title>
<link href="https://fonts.googleapis.com/css2?family=Roboto:wght@300;400;500;700&family=Roboto+Slab:wght@400;500;700&display=swap" rel="stylesheet">
<link href="https://fonts.googleapis.com/icon?family=Material+Icons" rel="stylesheet">
<style>
* {
margin: 0;
padding: 0;
box-sizing: border-box;
}
body {
font-family: 'Roboto', sans-serif;
background-color: #f8f9fa;
color: #1a237e;
line-height: 1.6;
}
.poster-container {
width: 920px;
min-height: 960px;
margin: 0 auto;
background: linear-gradient(135deg, #e8eaf6 0%, #c5cae9 100%);
padding: 40px;
position: relative;
overflow: hidden;
}
.poster-container::before {
content: "";
position: absolute;
top: -150px;
right: -150px;
width: 400px;
height: 400px;
border-radius: 50%;
background: linear-gradient(45deg, rgba(63, 81, 181, 0.1), rgba(103, 58, 183, 0.1));
z-index: 0;
}
.poster-container::after {
content: "";
position: absolute;
bottom: -100px;
left: -100px;
width: 300px;
height: 300px;
border-radius: 50%;
background: linear-gradient(45deg, rgba(63, 81, 181, 0.1), rgba(103, 58, 183, 0.1));
z-index: 0;
}
.header {
text-align: center;
margin-bottom: 30px;
position: relative;
z-index: 1;
}
.title {
font-family: 'Roboto Slab', serif;
font-size: 42px;
font-weight: 700;
color: #303f9f;
margin-bottom: 10px;
line-height: 1.2;
}
.subtitle {
font-size: 22px;
font-weight: 500;
color: #5c6bc0;
margin-bottom: 20px;
}
.section {
background-color: rgba(255, 255, 255, 0.85);
border-radius: 12px;
padding: 20px;
margin-bottom: 25px;
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.08);
position: relative;
z-index: 1;
}
.section-title {
font-family: 'Roboto Slab', serif;
font-size: 28px;
font-weight: 700;
color: #3949ab;
margin-bottom: 15px;
display: flex;
align-items: center;
}
.section-title .material-icons {
margin-right: 10px;
color: #5c6bc0;
}
.section-content {
font-size: 18px;
}
.highlight {
background: linear-gradient(transparent 40%, rgba(124, 77, 255, 0.2) 40%, rgba(124, 77, 255, 0.2) 85%, transparent 85%);
padding: 0 2px;
}
.problem-box {
background-color: rgba(239, 83, 80, 0.1);
border-left: 4px solid #ef5350;
padding: 15px;
margin: 15px 0;
border-radius: 0 8px 8px 0;
}
.solution-box {
background-color: rgba(76, 175, 80, 0.1);
border-left: 4px solid #4caf50;
padding: 15px;
margin: 15px 0;
border-radius: 0 8px 8px 0;
}
.architecture {
display: flex;
justify-content: space-between;
margin: 20px 0;
}
.component {
flex: 1;
background-color: rgba(63, 81, 181, 0.08);
border-radius: 8px;
padding: 15px;
margin: 0 5px;
text-align: center;
box-shadow: 0 2px 6px rgba(0, 0, 0, 0.05);
}
.component-title {
font-weight: 700;
color: #3949ab;
margin-bottom: 10px;
font-size: 20px;
}
.component-desc {
font-size: 16px;
}
.arrow {
display: flex;
align-items: center;
justify-content: center;
color: #5c6bc0;
}
.results-container {
display: flex;
justify-content: space-between;
margin: 20px 0;
}
.result-box {
flex: 1;
background-color: rgba(63, 81, 181, 0.08);
border-radius: 8px;
padding: 15px;
margin: 0 5px;
text-align: center;
}
.result-number {
font-size: 36px;
font-weight: 700;
color: #3949ab;
}
.result-label {
font-size: 16px;
}
.efficiency-table {
width: 100%;
border-collapse: collapse;
margin: 15px 0;
}
.efficiency-table th, .efficiency-table td {
padding: 10px;
text-align: left;
border-bottom: 1px solid #e0e0e0;
}
.efficiency-table th {
background-color: rgba(63, 81, 181, 0.1);
color: #3949ab;
}
.context-image {
width: 100%;
border-radius: 8px;
margin: 15px 0;
box-shadow: 0 4px 8px rgba(0, 0, 0, 0.1);
}
.footer {
text-align: center;
margin-top: 30px;
font-size: 14px;
color: #5c6bc0;
position: relative;
z-index: 1;
}
.bullet-list {
list-style-type: none;
padding-left: 0;
}
.bullet-list li {
position: relative;
padding-left: 25px;
margin-bottom: 8px;
}
.bullet-list li::before {
content: "•";
position: absolute;
left: 0;
color: #5c6bc0;
font-weight: bold;
}
</style>
</head>
<body>
<div class="poster-container">
<div class="header">
<h1 class="title">Agentic Context Engineering</h1>
<h2 class="subtitle">Evolving Contexts for Self-Improving Language Models</h2>
</div>
<div class="section">
<h3 class="section-title">
<i class="material-icons">info</i>
Introduction
</h3>
<div class="section-content">
<p>Large Language Model applications increasingly rely on <span class="highlight">context adaptation</span> rather than weight updates. Current approaches suffer from two critical limitations:</p>
<div class="problem-box">
<p><strong>Brevity bias:</strong> Over-prioritizing concise summaries at the expense of detailed domain insights</p>
<p><strong>Context collapse:</strong> Iterative rewriting erodes details over time, leading to performance drops</p>
</div>
<p>ACE treats contexts as <span class="highlight">evolving playbooks</span> that accumulate, refine, and organize strategies through a modular process.</p>
</div>
</div>
<div class="section">
<h3 class="section-title">
<i class="material-icons">architecture</i>
Three-Role Architecture
</h3>
<div class="section-content">
<div class="architecture">
<div class="component">
<div class="component-title">Generator</div>
<div class="component-desc">Produces reasoning trajectories for new queries, surfacing effective strategies and pitfalls</div>
</div>
<div class="arrow">
<i class="material-icons">arrow_forward</i>
</div>
<div class="component">
<div class="component-title">Reflector</div>
<div class="component-desc">Critiques generated traces, distilling insights from successes and errors</div>
</div>
<div class="arrow">
<i class="material-icons">arrow_forward</i>
</div>
<div class="component">
<div class="component-title">Curator</div>
<div class="component-desc">Synthesizes insights into structured "delta entries" and integrates them into existing context</div>
</div>
</div>
</div>
</div>
<div class="section">
<h3 class="section-title">
<i class="material-icons">lightbulb</i>
Key Innovations
</h3>
<div class="section-content">
<div class="solution-box">
<h4 style="color: #3949ab; margin-bottom: 10px;">Incremental Delta Updates</h4>
<ul class="bullet-list">
<li>Contexts represented as structured, itemized "bullets" with metadata and content</li>
<li>Small, localized edits preserve prior knowledge while accumulating new insights</li>
<li>Non-LLM logic for deterministic merging, de-duplication, and pruning</li>
</ul>
</div>
<div class="solution-box">
<h4 style="color: #3949ab; margin-bottom: 10px;">Grow-and-Refine Mechanism</h4>
<ul class="bullet-list">
<li>Balances context expansion with periodic refinement</li>
<li>Maintains relevance and prevents unbounded growth</li>
<li>Enables efficient, parallel merging crucial for scalability</li>
</ul>
</div>
</div>
</div>
<div class="section">
<h3 class="section-title">
<i class="material-icons">trending_up</i>
Performance Results
</h3>
<div class="section-content">
<p>ACE consistently outperforms strong baselines across agent and domain-specific benchmarks:</p>
<div class="results-container">
<div class="result-box">
<div class="result-number">+10.6%</div>
<div class="result-label">Agent Tasks (AppWorld)</div>
</div>
<div class="result-box">
<div class="result-number">+8.6%</div>
<div class="result-label">Financial Analysis (FiNER + XBRL)</div>
</div>
</div>
<p>Matches top-ranked production-level agent on AppWorld leaderboard using smaller open-source model.</p>
<img src="https://sfile.chatglm.cn/moeSlide/image/2d/2dc07c66.jpg" alt="Context-Quality Curve" class="context-image">
</div>
</div>
<div class="section">
<h3 class="section-title">
<i class="material-icons">speed</i>
Efficiency Gains
</h3>
<div class="section-content">
<p>ACE achieves significant efficiency improvements compared to existing methods:</p>
<table class="efficiency-table">
<tr>
<th>Metric</th>
<th>Offline vs GEPA</th>
<th>Online vs Dynamic Cheatsheet</th>
</tr>
<tr>
<td>Latency Reduction</td>
<td>82.3%</td>
<td>91.5%</td>
</tr>
<tr>
<td>Rollout/Token Cost Reduction</td>
<td>75.1%</td>
<td>83.6%</td>
</tr>
</table>
<p>Adapts effectively <span class="highlight">without labeled supervision</span> by leveraging natural execution feedback.</p>
</div>
</div>
<div class="section">
<h3 class="section-title">
<i class="material-icons">insights</i>
Implications
</h3>
<div class="section-content">
<ul class="bullet-list">
<li>Enables scalable, efficient, and self-improving LLM systems with low overhead</li>
<li>Provides interpretable contexts and lower overhead compared to fine-tuning</li>
<li>Offers a flexible approach for online and continuous learning</li>
<li>Particularly valuable for specialized domains and long-context applications</li>
</ul>
</div>
</div>
<div class="footer">
<p>arXiv:2510.04618 | Code available at github.com/ace-agent/ace</p>
</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！