Step Game TrueSkill Mu

Name: Step Game TrueSkill Mu
Published: 2025-06-06

Games · 2025-06-06

Lech Mazur's Multi-Agent Step Race benchmark evaluates models in a public-talk, private-move game where collisions prevent players from advancing. The public leaderboard reports TrueSkill mu as the headline measure of collaboration and deception under pressure.

Top models (higher is better)

Model	Score
GPT-5	5.5
o3	5.3
GPT-5.1	5.3
Gemini 3 Pro Preview	5.0
o1 (2024-12-17)	4.1
Gemini 2.5 Flash	4.0
Gemini 2.5 Pro	3.8
Grok 4.1 Fast	3.8
Gemini 2.5 Pro Preview 03-25	3.7
DeepSeek-V3.2-Exp	3.7
o3-mini	3.6
DeepSeek-V3.1	3.6
Grok 4	3.5
Sonnet 4.5	3.4
Kimi K2 Thinking	3.3
Grok 3 Mini	3.2
DeepSeek-R1-0528	3.2
QwQ-32B	3.2
DeepSeek-R1	3.2
Opus 4.5	3.2
Opus 4.1	3.1
Qwen3-235B-A22B	3.1
Sonnet 4	3.1
Silent Greedy (Lech Mazur baseline strategy)	3.0
GPT-5 Mini	2.8
Claude 3.7 Sonnet	2.8
O4 Mini	2.6
Silent Closer (Lech Mazur baseline strategy)	2.6
Silent EVMax (Lech Mazur baseline strategy)	2.6
Silent Collision Tracker (Lech Mazur baseline strategy)	2.5
Qwen3 235B A22B Instruct 2507	2.5
Silent Blocker (Lech Mazur baseline strategy)	2.4
o1-mini	2.4
Opus 4	2.4
Gemini 2.0 Flash Thinking Experimental 01-21	2.3
GLM 4.6	2.2
Qwen3-30B-A3B	2.2
GLM-4.5	1.7
Claude 3.5 Sonnet (Oct 2024)	1.7
gpt-oss-120b	1.6