🏆 Clash of Gods · Leaderboard

Here is the latest evaluation record under the academy_3_vs_1_with_keeper scenario.

The leaderboard data on this page is dynamically rendered from the combat records after fully automated practical matchups (max 400 physics steps + 5 step intervals) executed by the large language models via run_multiple_experiments.py.

排名	🤖 模型 (Model)	⚽ 进球胜率	🌟 平均分 (Reward)	⏳ 场均步数	⚡ 响应延迟	❌ 解析失败率	场次
🏆	GLM_5 (Mock)	80.0%	1.28	114.4	6611 ms	0.0%	5
🥈	Gemini_3_0_Flash (Mock)	40.0%	1.19	258.4	4611 ms	0.0%	5

🎖️ Match Commentary

Through the real-time perspective data above, we observed the following conclusions within the decision space of Large Language Models:

• Goal Terminator: GLM-5 demonstrated an outstanding 80% goal rate and the dominant capability of a lightning strike on average 114 steps per match.
• Speed of Thought: Gemini-3.0-Flash utilized its miniature size to secure extremely fast response latency (millisecond level).
• Absolute Zero Errors: Thanks to the excellent regex fallback parser system in our engineering design, the instruction parsing crash rate for all tested contestants was 0.0%.