Beyond Self-Interest: Modeling Social-Oriented Motivation
for Human-like Multi-Agent Interactions

AAMAS (Oral)
Jingzhe Lin1, Ceyao Zhang2,3, Yaodong Yang2, Yizhou Wang4, Song-Chun Zhu4,5, ✉ Fangwei Zhong1
1 School of Artificial Intelligence, Beijing Normal University, Beijing, China.
2 School of Intelligence Science and Technology, Peking University, Beijing, China.
3 Institute for Artificial Intelligence, Peking University, Beijing, China.
4 Peking University, Beijing, China.
5 Beijing Institute for General Artificial Intelligence (BIGAI), Beijing, China.
✉ Corresponding Author
Paper Appendix Code

Abstract

ASVO (Autonomous Social Value-Oriented agents) is a multi-agent simulation framework that combines Social Value Orientation (SVO) with value-driven LLM agents. Instead of fixed behavior rules, agents maintain structured social desires and update them through context-aware reflection during interaction. This enables adaptive behavior over time, including shifts between cooperation and competition, and supports more realistic, interpretable social simulations.

Method Workflow

ASVO method workflow

ASVO models adaptive social motivation through a structured psychological profile and dynamic SVO adaptation. At each step, agents perform Belief Update, Value Update, SVO Calculation, and Activity Generation, forming a closed feedback loop. This design enables the emergence of human-like cooperation, competition, and personality drift in multi-agent social systems.

Examples

ASVO agents collaborating in a workplace scenario ASVO agents presenting results in a workplace scenario

Simulated workplace scenarios with different personality traits

ASVO agents collaborating in a workplace scenario ASVO agents presenting results in a workplace scenario

Simulated family scenarios with different personality traits

ASVO agents collaborating in a workplace scenario ASVO agents presenting results in a workplace scenario

Simulated classroom scenarios with different personality traits

Result

LLM ReAct BabyAGI LLMob D2A ASVO
NH NH NH NH NH
School
Deepseek 4.0043.667 3.4382.771 4.5003.708 3.2923.083 4.7504.792
GPT-5 3.7503.417 3.8333.521 4.3544.271 3.8123.448 4.9584.958
Gemini-2.5 3.8333.896 3.4793.042 4.3124.062 3.4273.083 4.7084.708
Qwen3 3.9173.521 3.4583.188 4.2714.021 3.6773.406 4.7924.824
Avg 3.8763.625 3.5523.130 4.3593.938 3.5523.255 4.8024.821
Std 1.0970.658 1.3220.877 0.8610.565 1.4261.047 0.4970.456
Workplace
Deepseek 3.2142.385 2.4552.397 3.2532.609 2.8492.479 4.7664.119
GPT-5 3.2213.045 3.3493.064 4.1443.792 3.7033.458 4.8494.019
Gemini-2.5 3.4483.182 2.6732.724 3.4423.269 2.8912.698 4.8784.034
Qwen3 3.2972.760 2.8432.673 3.3303.058 3.1352.818 4.7824.026
Avg 3.3302.877 2.7982.710 3.5423.182 3.1452.863 4.8194.049
Std 1.5681.063 1.3580.968 1.4631.088 1.4440.991 0.4910.466
Family
Deepseek 3.2863.262 2.9172.690 3.5893.339 4.0103.167 4.7503.964
GPT-5 3.3453.202 3.3453.202 4.2023.863 4.4793.750 4.9054.024
Gemini-2.5 3.7263.488 3.4231.433 3.8453.429 4.0313.615 4.8634.077
Qwen3 4.0063.536 3.6853.315 4.0003.738 4.4903.781 4.3813.820
Avg 3.6933.449 3.3423.079 3.9093.579 4.2533.578 4.7253.946
Std 1.4230.921 1.2390.904 1.1660.827 1.1480.941 0.6240.648

Across school, workplace, and family settings, ASVO consistently attains the highest average scores and the lowest variance, indicating both stronger overall performance and more stable behavior compared to ReAct, BabyAGI, LLMob, and D2A baselines.

Figure 3: ASVO analysis

ASVO’s action distribution stays aligned with its configured SVO persona, cooperative under prosocial setups and increasingly competitive as the angle turns individualistic, whereas baselines with identical profiles produce mixed behaviors that stray from their intended orientations.

Citation

@inproceedings{asvo,
      author  = {Lin, Jingzhe and Zhang, Ceyao and Yang, Yaodong and Wang, Yizhou and Zhu, Song-Chun and Zhong, Fangwei},
      title   = {Beyond Self-Interest: Modeling Social-Oriented Motivation for Human-like Multi-Agent Interactions},
      booktitle = {Proceedings of the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026)},
      year    = {2026},
      address = {Paphos, Cyprus},
      publisher = {Association for Computing Machinery}
    }