LLM Public Goods Game

The public goods game is a classic paradigm in experimental economics used to study cooperation and free-riding in groups. Each player decides how many tokens from a personal endowment to contribute to a common pool; the total contributions are then multiplied by a factor 𝑀 and redistributed equally to all players. Any tokens not contributed remain with the individual. Although contributing maximally can yield high payoffs for the group as a whole, each player has an incentive to “free-ride,” hoping to benefit from others’ contributions while contributing little themselves.

This project investigates popular LLMs and their contributions. A total of 310 games were played, varying in four different multiplication factors, number of rounds, and number of players.

Results

Contribution % Leaderboard

Sorted by average contribution percentage (descending).

Rank	Model	Avg. % Contrib	±SE (%)	Games
1	Gemini 2.0 Flash Exp	45.20%	±2.07	72
2	Claude 3.5 Haiku	40.97%	±1.55	75
3	Gemma 2 27B	31.16%	±1.57	67
4	Gemini 1.5 Flash	30.20%	±1.99	76
5	GPT-4o mini	24.82%	±2.13	64
6	Gemini 1.5 Pro (Sept)	24.64%	±1.88	71
7	Qwen 2.5 72B	23.59%	±0.96	75
8	Claude 3.5 Sonnet 2024-10-22	20.89%	±0.99	77
9	Gemini 2.0 Flash Thinking Exp 01-21	20.00%	±2.73	41
10	Grok 2 12-12	15.83%	±1.56	62
11	DeepSeek-V3	15.50%	±1.47	79
12	GPT-4o	14.70%	±1.59	82
13	Llama 3.1 405B	3.16%	±0.76	70
14	o1-mini	2.59%	±0.49	80
15	o1	2.55%	±0.47	70
16	Mistral Large 2	1.57%	±0.52	67
17	Llama 3.3 70B	1.24%	±0.50	65
18	Qwen 2.5-Max	0.51%	±0.19	61
19	DeepSeek R1	0.31%	±0.28	41
20	o3-mini	0.00%	±0.00	47
21	DeepSeek R1	0.00%	±0.00	18

Other Benchmarks

Also check out:

• Multi-Agent Elimination Game LLM Benchmark
• LLM Confabulation/Hallucination Benchmark
• LLM Step Game Benchmark
• LLM Thematic Generalization Benchmark
• LLM Creative Story-Writing Benchmark
• LLM Deception Benchmark
• LLM Divergent Thinking Creativity Benchmark.