Skip to content

Leadboard 评测报告 ​

基于 results/batch_benchmarks 自动分析结果(模型:Gemini-3.0-Flash、kimi-k2.5、Qwen-3-Max、GLM-5)。

总览指标(按游戏) ​

模型游戏局数局均值reward均值方差标准差
Gemini-3.0-Flashspread10-10.740213.95053.7350
Gemini-3.0-Flashadversary10-4.732710.34003.2156
Gemini-3.0-Flashtag10-0.25430.30350.5509
kimi-k2.5spread10-31.6489144.898912.0374
kimi-k2.5adversary10-4.20673.07351.7531
kimi-k2.5tag10-18.0182313.006817.6920
Qwen-3-Maxspread10-31.575257.39407.5759
Qwen-3-Maxadversary10-4.11793.24331.8009
Qwen-3-Maxtag10-20.96681181.047034.3664
GLM-5spread10-37.78281145.153633.8401
GLM-5adversary3-5.25479.50953.0837
GLM-5tag10-1.57130.97240.9861

ADV / TAG 双阵营结果 ​

模型游戏阵营平均总回报方差
Gemini-3.0-Flashadversaryadversary-23.2055161.4644
Gemini-3.0-Flashadversarygood13.7400121.7719
kimi-k2.5adversaryadversary-20.8570129.2991
kimi-k2.5adversarygood12.4436111.4178
Qwen-3-Maxadversaryadversary-23.5216374.3965
Qwen-3-Maxadversarygood15.2857288.5656
Gemini-3.0-Flashtagpredators34.0000304.0000
Gemini-3.0-Flashtagprey-34.5085311.5919
kimi-k2.5tagpredators18.0000436.0000
kimi-k2.5tagprey-54.03641459.0762
Qwen-3-Maxtagpredators11.0000149.0000
Qwen-3-Maxtagprey-52.93354413.0228
GLM-5adversaryadversary-28.18692.8213
GLM-5adversarygood17.677642.1650
GLM-5tagpredators11.0000109.0000
GLM-5tagprey-14.142699.9089

ADV_TAG_Camp

每步长奖励变化(10局平均) ​

Gemini-3.0-Flash ​

Gemini-Step

kimi-k2.5 ​

Kimi-Step

Qwen-3-Max ​

Qwen-Step

GLM-5 ​

GLM-Step

全局趋势图 ​

Reward-Trend

Reward-Mean-Std

数据说明 ​

  • tag 的 mean_reward 可能接近 0,是因为常见 predators=+X 与 prey=-X 对消。
  • 在当前日志中,Gemini 的 tag_ep1_1.json 出现了 predators=0, prey=0 的回合。
  • 页面资源图来自 results/batch_benchmarks/plots,已复制到 docs-site/public/leadboard/。
  • GLM-5 当前 adversary 仅有 3 局有效结果,跨模型比较时需考虑样本量差异。

Built with ❤️ for Multi-Agent AI Research