Benchmark Atlas

亚洲模型实验室 · arXiv 论文

Capacity-aware Recycle Routing

复现 Hunyuan-Large Figure 2:同一个 overloaded token 在传统 top-k 中被丢弃,在 recycle routing 中被二次分配给未满专家。

hunyuanmoeroutingcapacitytoken-flow
Overloaded MoE tokens · drop vs recycle复现 Hunyuan-Large Figure 2:同一个 overloaded token 在传统 top-k 中被丢弃,在 recycle routing 中被二次分配给未满专家。 示意数据,仅复现信息结构,不代表最新榜单结果。 Overloaded MoE tokens · drop vs recycleHunyuan-Large Technical Report · Figure 2 / PDF p.6DEMO DATATRADITIONAL TOP-KABCDEFrouterE1ACE2E3E4BE5DROP DRECYCLE ROUTINGABCDEFrouterE1ACE2E3E4BDE5capacity-aware second passexpert capacity = 2 · overloaded token either disappears or is reassigned to an available expertstate transition preserves the individual token identity across both panels
Illustrative demo data · renderer: recycleRoutingComparison

Information grammar

它如何组织信息

传统/回收双路由面板 + token 身份保持 + expert 容量槽 + overloaded token 丢弃/二次分配状态转换

Visual system
VS-80 Hunyuan token-preserving recycle route
Component ID
hunyuan-recycle-routing

Use when

适用判断

适合解释 MoE 容量溢出策略;必须同时展示容量上限、token 身份和二次路由结果。

Figure-level evidence

Figure 2 · PDF p.6

Expert 1 容量为 2;Token D 从丢弃状态变为重新分配到未满的 Expert 4。

Verified
2026-07-13
Paper source
Hunyuan-Large 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。

{
  "methods": [
    {
      "label": "TRADITIONAL TOP-K",
      "tokens": [
        "A",
        "B",
        "C",
        "D",
        "E",
        "F"
      ],
      "experts": [
        [
          "A",
          "C"
        ],
        [],
        [],
        [
          "B"
        ],
        []
      ],
      "recycle": false
    },
    {
      "label": "RECYCLE ROUTING",
      "tokens": [
        "A",
        "B",
        "C",
        "D",
        "E",
        "F"
      ],
      "experts": [
        [
          "A",
          "C"
        ],
        [],
        [],
        [
          "B",
          "D"
        ],
        []
      ],
      "recycle": true
    }
  ]
}

Related components

同一图表家族的其他语法