{
  "source": "original paper only",
  "source_url": "https://arxiv.org/html/2610.13092v1",
  "local_run": false,
  "reviewed_at": "2026-10-12",
  "rows": [
    {
      "metric": "total executions",
      "reported_value": "4500",
      "comparator_or_protocol": "ALFWorld/AndroidWorld，各配置三次"
    },
    {
      "metric": "T3A cost per delivered success, base/Jev",
      "reported_value": "68.70 / 73.97 mUSD",
      "comparator_or_protocol": "+7.7%，95%区间跨0"
    },
    {
      "metric": "FastAgent time per delivered success, base/Jev",
      "reported_value": "239.4 / 377.8 seconds",
      "comparator_or_protocol": "成功率下降削弱单步收益"
    }
  ]
}