亚洲模型实验室 · arXiv 论文
Matched-quality RL Speedup
复现 MiniMax-M1 Figure 2：除最终分数外，直接比较 CISPO 与 DAPO 达到相同性能所需的训练步数。
matchedQualitySpeedupInformation grammar
它如何组织信息
多算法训练轨迹 + 同分水平双向步数箭头 + 等质量 speedup 直标
- Visual system
- VS-68 MiniMax matched-quality speed arrow
- Component ID
minimax-m1-matched-quality-speedup
Use when
适用判断
适合算法训练效率对比；匹配点必须来自同一任务、模型底座、采样与评测协议。
Figure-level evidence
Figure 2 · PDF p.6
GRPO、DAPO、CISPO 的 AIME avg@32 训练轨迹及 2x matched-quality speedup。
- Verified
- 2026-07-12
- Paper source
- MiniMax-M1 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。
{
"steps": [
0,
100,
200,
300,
400,
500,
600,
700,
800,
900,
1000,
1100,
1200,
1300,
1400,
1500
],
"series": [
{
"label": "GRPO",
"values": [
2,
10,
17,
20,
22,
21,
21,
22,
23,
22,
23,
24,
23,
24,
22,
23
]
},
{
"label": "DAPO",
"values": [
2,
8,
15,
21,
21,
26,
24,
27,
29,
28,
31,
28,
32,
28,
33,
34
]
},
{
"label": "CISPO",
"values": [
2,
9,
24,
26,
28,
33,
35,
36,
38,
42,
41,
44,
40,
42,
41,
39
]
}
],
"match": {
"score": 33,
"candidateStep": 500,
"referenceStep": 1000,
"speedup": 2
}
}Related components