亚洲模型实验室 · arXiv 论文
Loss-to-Benchmark Transfer Calibration
复现 Seed1.5-VL Figure 3：把 OCR/grounding 子类 loss 的 token 缩放与相关下游评测指标的局部校准关系串成两段证据链。
lossMetricTransferFacetsInformation grammar
它如何组织信息
能力子类分行 + token-loss 幂律首面板 + 同一 loss 映射多个下游指标的局部 log-linear 双面板 + 因果阅读箭头
- Visual system
- VS-81 Seed linked loss-transfer facets
- Component ID
seed-loss-metric-transfer
Use when
适用判断
适合用训练代理指标预测相关 benchmark；只应在论文支持的局部 loss 区间内解释，不外推到饱和区。
Figure-level evidence
Figure 3 · PDF p.14
六分面分别连接 OCR/grounding loss 的 token 幂律与 ChartQA、InfoVQA、RefCOCO、RefCOCO+ 指标。
- Verified
- 2026-07-13
- Paper source
- Seed1.5-VL 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。
{
"rows": [
{
"label": "OCR",
"panels": [
{
"label": "loss vs training tokens",
"axis": "log tokens →",
"points": [
[
0.04,
0.92
],
[
0.15,
0.82
],
[
0.28,
0.77
],
[
0.4,
0.64
],
[
0.54,
0.55
],
[
0.65,
0.43
],
[
0.76,
0.32
],
[
0.9,
0.16
],
[
0.98,
0.08
]
]
},
{
"label": "ChartQA vs -log loss",
"axis": "-log OCR loss →",
"points": [
[
0.08,
0.12
],
[
0.2,
0.24
],
[
0.32,
0.3
],
[
0.46,
0.48
],
[
0.62,
0.55
],
[
0.76,
0.71
],
[
0.91,
0.84
]
]
},
{
"label": "InfoVQA vs -log loss",
"axis": "-log OCR loss →",
"points": [
[
0.06,
0.1
],
[
0.22,
0.2
],
[
0.38,
0.36
],
[
0.54,
0.47
],
[
0.68,
0.64
],
[
0.82,
0.78
],
[
0.95,
0.86
]
]
}
]
},
{
"label": "GROUNDING",
"panels": [
{
"label": "loss vs training tokens",
"axis": "log tokens →",
"points": [
[
0.02,
0.94
],
[
0.17,
0.85
],
[
0.31,
0.71
],
[
0.46,
0.62
],
[
0.6,
0.48
],
[
0.73,
0.37
],
[
0.86,
0.23
],
[
0.98,
0.1
]
]
},
{
"label": "RefCOCO vs -log loss",
"axis": "-log grounding loss →",
"points": [
[
0.08,
0.16
],
[
0.22,
0.28
],
[
0.36,
0.25
],
[
0.49,
0.52
],
[
0.63,
0.58
],
[
0.77,
0.73
],
[
0.93,
0.9
]
]
},
{
"label": "RefCOCO+ vs -log loss",
"axis": "-log grounding loss →",
"points": [
[
0.07,
0.12
],
[
0.22,
0.3
],
[
0.37,
0.36
],
[
0.51,
0.49
],
[
0.66,
0.64
],
[
0.8,
0.76
],
[
0.95,
0.92
]
]
}
]
}
]
}Related components