亚洲模型实验室 · arXiv 论文
Cross-modal Expert Collaboration Facets
合并复现 ERNIE 5.0 Figures 8-9：用首层、中层、末层的专家条码和 IoU 摘要观察专家专门化与跨模态协作。
expertCollaborationFacetsInformation grammar
它如何组织信息
首中末层列分面 + 模态行 expert 条码 + 底部跨模态 top-expert IoU 摘要
- Visual system
- VS-75 ERNIE expert collaboration facets
- Component ID
ernie-expert-collaboration
Use when
适用判断
适合 MoE routing 诊断；每行需要相同 expert ID 顺序，IoU 的 top-k 口径必须固定。
Figure-level evidence
Figures 8-9 · PDF p.24
六种模态/任务在首中末层的 expert utilization，以及 top 25% experts 的跨模态 IoU。
- Verified
- 2026-07-13
- Paper source
- ERNIE 5.0 Technical Report
Structured data
可替换的数据模型
示意数据，仅复现信息结构，不代表最新榜单结果。
{
"layers": [
{
"label": "FIRST LAYER",
"rows": [
{
"label": "Text",
"values": [
2,
4,
3,
1,
2,
3
]
},
{
"label": "Image U",
"values": [
11,
3,
18,
2,
5,
1
]
},
{
"label": "Video U",
"values": [
24,
4,
12,
1,
6,
2
]
},
{
"label": "Image G",
"values": [
2,
18,
4,
1,
8,
3
]
},
{
"label": "Video G",
"values": [
1,
3,
2,
22,
7,
4
]
},
{
"label": "Audio",
"values": [
1,
12,
3,
29,
4,
20
]
}
]
},
{
"label": "MIDDLE LAYER",
"rows": [
{
"label": "Text",
"values": [
3,
6,
2,
3,
4,
1
]
},
{
"label": "Image U",
"values": [
4,
2,
6,
3,
19,
2
]
},
{
"label": "Video U",
"values": [
8,
2,
5,
3,
7,
8
]
},
{
"label": "Image G",
"values": [
2,
14,
3,
7,
24,
4
]
},
{
"label": "Video G",
"values": [
1,
9,
2,
16,
40,
3
]
},
{
"label": "Audio",
"values": [
65,
20,
4,
2,
3,
5
]
}
]
},
{
"label": "LAST LAYER",
"rows": [
{
"label": "Text",
"values": [
2,
3,
13,
2,
5,
1
]
},
{
"label": "Image U",
"values": [
2,
5,
42,
3,
7,
1
]
},
{
"label": "Video U",
"values": [
2,
15,
27,
3,
7,
2
]
},
{
"label": "Image G",
"values": [
8,
2,
31,
22,
15,
3
]
},
{
"label": "Video G",
"values": [
15,
2,
26,
35,
5,
40
]
},
{
"label": "Audio",
"values": [
3,
14,
30,
7,
4,
2
]
}
]
}
],
"overlap": [
{
"label": "image ↔ video understanding",
"value": 0.64
},
{
"label": "image ↔ video generation",
"value": 0.78
},
{
"label": "text ↔ audio",
"value": 0.4
}
]
}Related components