ALSO: Adversarial Online Strategy Optimization for Social Agents
ALSO (Adversarial Online Strategy Optimization) is a new framework for online strategy optimization in multi-agent social simulation, addressing the limitations of static personas in LLM-based agents. It formulates multi-turn interaction as a dynamic, non-stationary environment where agents must adapt strategies over time. The framework introduces adversarial online learning to optimize strategies without offline training or external planners, reducing overhead. This approach is designed for social intelligence testbeds involving evolving contexts and strategically adapting opponents.
Key facts
- ALSO stands for Adversarial Online Strategy Optimization
- It is the first framework for online strategy optimization in multi-agent social simulation
- Addresses non-stationary environments in social simulation
- LLM-based social agents typically rely on static personas
- Existing approaches like offline RL or external planners assume stationarity
- ALSO uses adversarial online learning for strategy adaptation
- Reduces training overhead compared to offline methods
- Formulates multi-turn interaction as a dynamic process
Entities
—