智能体进化新刻度：字节Seed发布EdgeBench基准测试
<div class="ce-block ce-block--focused" data-id="dAtbMsKC7U" ><div class="ce-block__content">
<div class="ce-paragraph cdx-block ce-paragraph--left">在人工智能技术快速演进的今天，如何科学地衡量智能体（Agent）在真实世界中的持续学习能力，成为了学术界与工业界共同关注的焦点。近日，字节Seed团队正式发布了名为“EdgeBench”的超长程评测集，为这一领域的研究提供了全新的量化参考。</div>
</div>
</div><div class="ce-block ce-block--focused" data-id="aP4TKqR7pM" >
<div class="ce-block__content">
<div class="ce-paragraph cdx-block ce-paragraph--left">EdgeBench的核心价值在于其对“真实世界环境学习”的深度覆盖。该基准收录了134个涵盖六大领域的真实任务，且每个任务都要求智能体能够持续工作至少12小时。这一设计旨在突破以往短时任务评测的局限，更真实地模拟智能体在复杂、动态环境下的长期表现。为了构建这一严谨的测试体系，研发团队累计采集了约3.8万小时的交互数据。</div>
</div>
</div><div class="ce-block ce-block--focused" data-id="9EnutqJTao" >
<div class="ce-block__content">
<div class="cdx-block image-tool image-tool--filled ">
<div class="image-tool__image">
<div class="image-tool__image-preloader" style=""></div>
<img id="aimg_9EnutqJTao" class="image-tool__image-picture _zoom" src="https://wazhi.net/data/attachment/forum/202607/07/190547qu0zfo0owozmqffm.png" title="" alt="" data-aid="240"/>
</div>
<div class="cdx-input image-tool__caption" data-placeholder=""></div>
</div>
</div>
</div><div class="ce-block ce-block--focused" data-id="L1VP6-ftml" >
<div class="ce-block__content">
<div class="ce-paragraph cdx-block ce-paragraph--left">研究结果揭示了一项值得关注的趋势:智能体在环境学习中的表现遵循着一条高精度的log-sigmoid曲线，其拟合优度（R²）高达0.998。这意味着智能体的学习过程呈现出极强的规律性。此外，数据分析显示，从2025年9月到2026年5月期间，前沿模型的学习速度展现出每三个月翻一番的强劲增长势头。</div>
</div>
</div><div class="ce-block ce-block--focused" data-id="QwCMH2wLGh" >
<div class="ce-block__content">
<div class="ce-paragraph cdx-block ce-paragraph--left">目前，EdgeBench已向开发者社区开源了其中的51个任务及配套的完整评测框架。虽然该基准目前仍处于学术探索阶段，但它<span class="spamTxt">首次</span>将长程环境学习规律进行了量化描述。对于AI研究者而言，这一规律不仅提供了衡量模型能力的硬指标，也为未来提升智能体的环境适应性和学习效率指明了方向。</div>
</div>
</div><style type="text/css">
.ce-block {
margin-bottom: 20px;
}
.ce-block__content,.ce-toolbar__content {
/* max-width:calc(100% - 50px) */
margin-left: auto;
margin-right: auto;
}
.ce-paragraph {
line-height: 1.6em;
outline: none;
text-indent: 2em;
font-size: 16px;
}
.ce-paragraph--right {
text-align: right;
}
.ce-paragraph--center {
text-align: center;
}
.ce-paragraph--left {
text-align: left;
}
.ce-paragraph--justify {
text-align: justify;
}
.ce-paragraph-text-indent {
text-align: justify;
}
.ce-paragraph:empty::before{
content: attr(data-placeholder);
color: #707684;
font-weight: normal;
opacity: 0;
}
/** Show placeholder at the first paragraph if Editor is empty */
.codex-editor--empty .ce-block:first-child .ce-paragraph:empty::before {
opacity: 1;
}
.codex-editor--toolbox-opened .ce-block:first-child .ce-paragraph:empty::before,
.codex-editor--empty .ce-block:first-child .ce-paragraph:empty:focus::before {
opacity: 0;
}
.ce-paragraph p:first-of-type{
margin-top: 0;
}
.ce-paragraph p:last-of-type{
margin-bottom: 0;
}
.svg-icon {
width: 1em;
height: 1em;
}
.svg-icon path,
.svg-icon polygon,
.svg-icon rect {
fill: #4691f6;
}
.svg-icon circle {
stroke: #4691f6;
stroke-width: 1;
}
</style><style type="text/css">
.ce-block {
margin-bottom: 20px;
}
.ce-block__content,.ce-toolbar__content {
/* max-width:calc(100% - 50px) */
margin-left: auto;
margin-right: auto;
}
.image-tool {
--bg-color: #cdd1e0;
--front-color: #388ae5;
--border-color: #e8e8eb;
}
.image-tool__image {
border-radius: 3px;
overflow: hidden;
margin-bottom: 10px;
}
.image-tool__image-picture {
max-width: 100%;
vertical-align: bottom;
display: block;
margin-left: auto;
margin-right: auto;
}
.image-tool__image-preloader {
width: 50px;
height: 50px;
border-radius: 50%;
background-size: cover;
margin: auto;
position: relative;
background-color: #cdd1e0;
background-position: center center;
}
.image-tool__image-preloader::after {
content: "";
position: absolute;
z-index: 3;
width: 60px;
height: 60px;
border-radius: 50%;
border: 2px solid #cdd1e0;
border-top-color: #388ae5;
left: 50%;
top: 50%;
margin-top: -30px;
margin-left: -30px;
animation: image-preloader-spin 2s infinite linear;
box-sizing: border-box;
}
.image-tool__caption::before {
position: absolute !important;
content: attr(data-placeholder);
color: #707684;
font-weight: normal;
display: none;
}
.image-tool__caption:empty::before {
display: block;
}
.image-tool__caption:empty:focus::before {
display: none;
}
.image-tool--empty .image-tool__image {
display: none;
}
.image-tool--empty .image-tool__caption, .image-tool--loading .image-tool__caption {
display: none;
}
.image-tool .cdx-button {
display: flex;
align-items: center;
justify-content: center;
}
.image-tool .cdx-button svg {
height: auto;
margin: 0 6px 0 0;
}
.image-tool--filled .cdx-button {
display: none;
}
.image-tool--filled .image-tool__image-preloader {
display: none;
}
.image-tool--loading .image-tool__image {
min-height: 200px;
display: flex;
border: 1px solid #e8e8eb;
background-color: #fff;
}
.image-tool--loading .image-tool__image-picture {
display: none;
}
.image-tool--loading .cdx-button {
display: none;
}
/**
* Tunes
* ----------------
*/
.image-tool--withBorder .image-tool__image {
border: 1px solid #e8e8eb;
}
.image-tool--withBackground .image-tool__image {
padding: 15px;
background: #cdd1e0;
}
.image-tool--withBackground .image-tool__image-picture {
max-width: 60%;
margin: 0 auto;
}
.image-tool--stretched .image-tool__image-picture {
width: 100%;
}
.image-tool__caption {
text-align: center;
font-size: 14px;
color: #a3a3a3;
}
@keyframes image-preloader-spin {
0% {
transform: rotate(0deg);
}
100% {
transform: rotate(360deg);
}
}
</style><style type="text/css">
.inline-code {
background: rgba(250, 239, 240, 0.78);
color: #b44437;
padding: 3px 4px;
border-radius: 5px;
margin: 0 1px;
font-family: inherit;
font-size: 0.86em;
font-weight: 500;
letter-spacing: 0.3px;
}
</style>
页:
[1]